Opus is still best in class for this, but it's worth noting how well Gemini 3.7 does vs a more comparable LLM price wise, which is Grok 4.6: https://html.non.io/neonRamenGrok4.6 . I thought Gemini would blow Grok out of the water (it generally has in the past), but Grok has really caught up.
Other thoughts: I really think Google has fallen behind here. Even as a high speed offering (this build took ~7min, which is pretty good!), it wont be able to claim dominance for long with cerebras announcing the Sol preview today: https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultraf... .
It's not a bad model by any means, but I just don't know what situation I'd reach for 3.7 Flash first for. Google really needs a differentiator, especially given how hard it is to get an API key from them. They can't be high friction and non-pareto.
Can you help me understand how it is hard to get an API key from Google? You just head on over to http://aistudio.google.com/api-keys and create a key... not any different from platform.openai.com?
Disclaimer: I work in Google so it might be that this link is not publicly well known
At a high level though, as a rule of thumb Google assumes that they're serving companies at Google scale first, and at a human scale second. For other companies it's the opposite. Generally what that means is the first experience you get with a Google product will route you through 8 different dashboards to set up ACLs before you've hired your 2nd employee.
FWIW I definitely did not not need to do anything like that to generate a key via AI Studio. It was like three clicks to get the free tier key, later on enabling billing was a few more plus typing in credit card info.
That is easy but I’ve also found myself in account setup dashboards that were obviously geared toward enterprise trying to set up access to tinker with something AI related. It might have been TTS but it’s been a little while and I can’t quite remember.
it worked for me okay when I needed it for myself in my personal account. But when I tried setup this for a company I spent almost a day solving lot of small puzzles in GCE like how to tell CEO that he have to connect billing account created for other purposes (and he not even remember at time that it exist) to new project and all other things that others talking about.
You need to create a Google Cloud project to create an api key and when you try to create one you very often get error messages like:
“Failed to create project, The request is suspicious. Please try again” or “ You do not have permission to create a key in this project”. You can then navigate multiple screens in GCP to make it work but it’s a hassle compared to any other provider (OAI/Ant/OpenRouter or any of the Chinese labs).
I didn't have that issue back when I originally created my API keys a couple years ago. Just out of curiosity I switched to a different Google account that had never interacted with AI Studio and never used Google Cloud console.
It was literally two clicks, and didn't even leave the page: the dialog asked to create a project and type in a name, I did that, clicked submit and then it was selected as the default project. One more click and I had the free tier API key.
Not saying you didn't have that experience at the time, but personally I have had zero issues with AI Studio and consider it the most simple/fastest dev dashboard to get started compared to the others like OpenAI/Anthropic (thanks to Google's free tier that lets you skip billing setup annoyances just to play around with Gemini).
Despite what HN threads (that are also frequently confused and talking about GCP instead) portray as universal/widespread issues or the process being complex and time consuming somehow.
I use LLMs rarely, and only for digging into subjects which I can't find enough information using search engines. I only tried Claude and Gemini, but Gemini both returns faster and higher quality information which I can use for more targeted digging myself.
Google being Google, their models tend to be better at finding, organizing and presenting information, from my experience.
Moving from either frontier intelligence or frontier latency to a single model that does both at the same time is potentially a game changer in certain industries. I can easily see e.g. hedge funds dropping tons of money on this, because it means they can now do the same thing as their competitors, but much faster. That's basically a license to print money.
I like using 3.5-flash-lite for doing cheap PDF and Image data extraction stuff. I don't think there is a better bang / buck model right now (3.1 is cheaper but a lot worse).
I think both outputs are really good. I don't see a lot of differences. So what exactly should be looking at and notice that one model did worse or better than the other one.
EDIT: OKAY I see it's mostly the "image" generation, not so much the HTML... Noticeable in the food photos and the foodtruck/cart photo
How are you doing this with Opus. Clearly I’m missing something. I always turn to ChatGPT when I need images because Opus typically refuses. I’ve tried Claude Code and Claude online in the past. I’m pretty sure neither created images for me and I thought this was because Anthropic was focused on code.
I have a few custom ones (a post-trained flux 2 checkpoint for web design and a image->metalness map generator that the build step can call for more advanced lighting situations), but gpt-image-2 is better than my own for design, so it's weighted much more heavily in outputs my tool generates. I think gpt-image-2 currently generates 99%+ of the design outputs on diffui
I believe they are testing giving it an image, which you can do in Claude code by dragging/dropping into the terminal or copy/pasting, and asking it to build the html equivalent.
There is some irony being a developer and reading along the lines of: "oh look at the comparison between these models executing a task for a few cents on a job i'd be charging 1k minimum"
I don't think I'd say rarely. Companies rarely allocate the design resources to produce that, but the companies that do are typically much larger, so the actual number of individual developers that get rich mocks is probably closer to 40-50%.
I'm curious how much the harness plays into this. I'm somewhat surprised by the gemini and grok results, they seem to have strongly deviated from the original images. I'm thinking maybe the harness has a big effect? It's possible to proxy in different models to claude code, if you're curious you might find it interesting to test!
I'm not sure what prompt you put in but did Gemini replace the all of the images in the original with its own? That would be really weird behavior unprompted.
The agent is told to generate assets as part of the buildout. It gets to decide what the prompt is for them / whether to do postprocessing like background removal / what type of asset to generate.
The "introductory pricing" for this 3.7 Flash model is really weird.
It's scheduled to double in price on December 31, 2026, but who would anticipate still using this model five months from now? Especially since 3.6 Flash came out just three weeks ago!
Then I ran it on high, medium and low thinking levels (oddly minimal is no longer an option, which WAS an option for 3.5 and 3.6) and got a pretty excellent pelican for the first two:
UPDATE: That was in Safari, but as pointed out in the replies here the pelicans do NOT render well in Firefox or Chrome! Best guess is that's because of this invalid filter in the SVG:
Filters are meant to contain additional elements, not be empty: https://drafts.csswg.org/filter-effects/#FilterElement - so maybe Chrome and Firefox remove the element that references the broken filter but Safari doesn't?
At $work we still have some places using Opus 4.5. Even places using Qwen 2.5-VL, which is now 18 months old. It works, and upgrading is work (we'd have to validate the new model performs comparable in all the corner cases that currently work just fine)
Those kind of workloads would be hit by an end of introductory pricing. And it's exactly the kind of cases that are not very price sensitive. Where we are price sensitive we track new model releases closely, where we aren't other issues get priority as long as llm performance is good enough
Ever since the insane discount with GPT-5.6 Luna, not much excites me anymore. I mean just look at the benchmarks, even though Gemini 3.7 Flash performs well on the DeepSWE 1.1, Luna (Max) still performs way better. I personally have stuck to Luna (Xhigh) because its been more than enough and does not bloat up the context window too fast with reasoning tokens.
GPT-5.6 Luna is an insanely powerful model for its price. It's been great for coding workflows where I guide the LLM's hand step by step. It's also insane to see my weekly limit drop by than 2% after an hour of coding ever since the discount.
However, I've noticed 2 drawbacks with Luna. Context rot is much more palpable than Terra and Sol. It tends to get confused and go into rabbit holes when it's context gets filled up. In addition, when instructions are vague, it performs poorly and tends to write way to more code than necessary, but that is to be expected of smaller models. In all, for clearly defined, bite-sized coding tasks, Luna's price-to-performance has been insane. It might have very well commanded the price tag of Sol if it came out just a year ago.
luna is the first model that has outdone gpt-5-mini on the pareto frontier for some of my high value, cost sensitive ai product workflows. it's both cheaper (by about 60% in real world use) and higher quality based on my test harnesses. I was really worried that costs would go up since there wasn't a replacement as of a few weeks ago and gpt-5-mini is scheduled to be sunset toward the end of the year. So long as they don't randomly sunset this model anytime soon, that worry has now subsided.
I am in the same boat as you. I am using Luna and DeepSeek Flash. Both super fast, super cheap, and I have not felt need for anything more capable in few weeks.
I'm curious how much people are manually curating context these days; I'm increasingly feeling for myself that it being auto-managed inside a front-end like claude code is not ideal, and I'd rather have more control over what exact files and pieces of discovery go into a particular prompt, and the ability to more easily "fork" a session and ask asides or make notes/todos in a way that doesn't disrupt or confuse a more focused task going on.
I don't think I want a gastown-style "just yolo everything" approach, in fact I really want more control over how decisions are made and with what info. Does this exist?
I am not sure this enlightens you with anything but I have a TODO.md file with three headlines. Todo, Doing and Done. The agent is aware of it and knows on which task we are on.
On complete, it moves the user story from Doing to Done. I also have a MEMORY.md file that the agent read and writes in the beginning of a new conversation and at the end of our conversation to update stale information. These files are referred to every time I start a new conversation.
Regarding forking, I know Codex has such button underneath each message that lets you fork the whole conversation. I usually do that when I want to sidetrack and discuss something.
I use no SKILLS or commands like /goal. I’ve come a long way with just prompts and markdown files. Its all a different way of encapsulating instructions anyways.
what do you mean by reset at every turn? context stays until compaction. if you remove the reasoning tokens after every turn you will be constantly blowing cache which is far worse than filling up context.
Does it reset at every turn? From my experience in Codex for example, Luna (Max) fills the 256k token window relatively quick. The only thing lowering the context window again is the compaction.
It’s a big difference but Luna is very usable. I’ve plugged it into the slot I used to have GLM 5.2 in; I think it’s just as good. And it is less costly. I have Sol do planning and design but do most task execution with Luna now.
They need to release benchmarks against Luna/Terra. Luna is much cheaper which feels like it undercuts the need for Flash.
I've always considered the Flash series of models to be for low-cost, high-volume, mostly text-based use cases (e.g. summarization, parsing, formatting), emphasis on low-cost.
flash-lite is more of their luna tier competitor but even still not quite there yet, but gemini's dominance on multimodal and image understanding i think really gets downplayed on this site when most people think the only think you can do with LLMs is write code
That's been my association as well. I see Flash get brought up a lot in relation to things like OCR and PDF processing frequently, and a lot of other routine multimodal workloads.
Yes this is my impression as well. To be fair I didn't compare to Luna yet, but Gemini 3.5 Lite is a very good and cheap multi-modal data extraction model.
Yes, was going to say I use it exclusively for video and audio. The ability to give it a YouTube link through the API and ask questions about it is awesome
So it's better than 3.6 Flash, at half the price. I've been pretty excited about Gemini models recently, they just feel so fast after spending most of the day at work waiting for Opus 5.
Yes, 3.6 Flash is very fast. I used to get a fair amount of usage of the Gemini Flash models on the free tier. I signed up for their $4.99/month tier (includes 400GB of Google space which was also enticing) and it turns out I only get about 15 to 20 minutes of usage before I get a come-back-in-7-days message. Comically low usage limits on that plan.
The Gemini Flash models makes perfect sense to me coming from a company like Google. Google AI Mode for search is a product I really find useful. It makes sense that Google focused on smaller, faster yet smart enough models that wouldn't break your bank on inference. It plays well into their product ecosystem.
Google AI Mode consistently gets me consistently good results and good speeds. It really changes what "googling" is for me.
I like Gemini (I'm just a dumb person without knowledge of 'benchmarks' or how x compares to y) while understanding that shoving it into Google auto summaries has been a bad idea and produces inaccurate results
It's funny that they don't mention this at all in the marketing or tech specs when it's obviously the biggest selling point by far. Without this it would be completely irrelevant.
Worth noting that OpenAI just announced that they got the full GPT 5.6 Sol model running on Cerebras at 750 tokens per second. No announcement of the pricing though...
> It's funny that they don't mention this at all in the marketing or tech specs when it's obviously the biggest selling point by far. Without this it would be completely irrelevant.
Good catch! You're right to point that out. My previous marketing copy missed that specific detail. Thank you for bringing it up!
Cerebras is crazy to watch on GPT OSS or Gemma, I feel like we need a new VibeOS demo but with Cerebras, the OS would literally build itself in a few seconds.
I use this in a customer facing application and Gemini’s speed makes the experience feel much better.
The application isn’t so complicated that you need opus level reasoning or code writing, we need “good enough” data retrieval and processing with natural language queries and the ability to answer follow up questions.
I've blown away by flash 3.6's speed while Opus chugs along for _hours_ on similar tasks. I've gotten into a opus designed -> gemini implemented -> opus reviewed dev cycle recently.
I am actively using Gemini flash to "translate" what Opus says into human language. I let opus do the design (with my assistance) and implementation, but then the report that Opus writes gets translated by Gemini so that I don't have to waste time to understand it.
Sol high is almost the same speed if you take into account drastically lower token use. Look at the artificial analysis speed vs token use. Gemini is 7x faster but 5x more tokens. And that's with Sol high being a substantially better model.
Edit: and Sol medium actually has the same AA intelligence score as Gemini 3.7, and has >7x fewer tokens, actually making it faster
is presumes you are doing longer difficult agentic tasks, if youre doing a simple problem in 1 or 2 shots, not really multi turn then theres no comparison.
The multimodal abilities are great, but if you deal with text only, what is the benefit of using this over DS V4 Flash/Pro? 13-26x cheaper with comparable intelligence, and available across many different inference providers.
I fail to see the usecase where DS V4 Pro is not enough, but Flash 3.7 is - except multimodal.
Luna is similar, and also 8x cheaper. Source: artificialanalysis
The only benefit I can see is the speed, that looks to be outstanding, probably thanks to their TPUs.
for non-coding applications, i think speed is a real differentiator. Im building an app that uses LLMs for some functionality that the user would not have any reason to expect is using AI and therefore having then wait seconds or minutes is just not feasible. latency is a huge upside for me
Anything interacting with the real world seems like latency would be hugely important. Something more asynchronous friendly (like coding) is for obvious reasons over represented here
> 13-26x cheaper with comparable intelligence, and available across many different inference providers.
Well, compared to 2 months ago, it's no longer 100x more expensive for similar levels of quality...
If they continue monthly-ish releases by 3.9 - by Halloween - they should be close to the best in terms of what you get for what you pay for.
In 2 months, they've gone from basically the bottom of the pack to at least being somewhat usable and competitive.
OpenAI and Anthropic release in a month, and change things. OpenAI is claiming to be close to an Astra release - but that seems like a Fable type release - where they're just releasing a better more expensive model, not more cost effective models.
did you try to ingest 1M documents per hour with any provider except GCP with Flash? None work at scale. Deepseek, Luna, Mistral all fail. 1 in 3 requests is a fail. I stopped trying.
The only thing that works at scale is gemini flash.
I guess the question then becomes "are you sure you'll do text only?"
I could probably do text only for my workflow (feature development/debugging for web microservices) but sometimes it is easier to just toss a screenshot into the Claude prompt, so that gives it an edge.
If your workflow is 100%, certifiably never ever going to involve an image, then yeah, this isn't going to be huge.
Yeah we use it for auto-triage of incidents, attempts to auto-remediate, and escalation to human. But for actual development, it’s not a viable option for us.
Yup. I use it for a ton of mundane queries (stuff that I might have used Google search for in the past) and it's great. Nice and fast and correct more often than not, especially if you prompt it in a way that it invokes Google search (but filters out ads and SEO slop). It's even alright at programming tasks but if it stumbles then I'll escalate to Gemini Pro with extended thinking.
> Coding and agentic tasks: Significantly higher quality on real-world software engineering and agentic benchmarks, improving issue resolution and reducing failed agent loops.
> Web development and stronger design parity: Generates higher-fidelity desktop and web application code directly from design mocks, with strong gains in design adherence and in auditing existing codebases against mocks to verify 1:1 design parity.
> Promotional pricing: Gemini 3.7 Flash will be available at an introductory price of $0.75/1M input tokens and $3.75/1M output tokens. We’re also applying this new rate to 3.6 Flash. Introductory pricing expires on December 31, 2026; after, $1.50/1M input tokens and $7.50/1M output tokens will apply.
Still no sign of 3.5 Pro. Will have to test it, low expectations given every other model from the Gemini 3 lineage, but one can hope. Just struggle to understand the promotional pricing being temporary for four months. Given this industry, I'd be hard pressed if 3.7 Flash was still in use by end of year, so why not make it the official pricing?
>Given this industry, I'd be hard pressed if 3.7 Flash was still in use by end of year, so why not make it the official pricing
It was probably to placate some kind of general internal pricing/revenue benchmark that doesn't account for new model releases. Politicians do shit like this incessantly and it reeks of bureaucracy.
I suspect it's a bit of a signal to investors etc.
"Hey, we are not in a race to the bottom. This is our usual pricing, but this now is a promotion because we know we're coming from behind and need to entice users."
They're drawing a line in the sand on monetisation and signalling that to everyone, while in reality offering it a deep discount (no idea if profitable or not) knowing that this model will probably be obsolete before then.
I don't get it, Google could heavily subsidy their Gemini models to make it more attractive, but they prefer to not do it. I don't know one soul who is using Gemini models to code.
Even OpenAI who doesn't have money or capacity is offering their Luna model at $1.2 per 1M/out.
Why would they? Unless they have lots of unused tpu real estate that they could host it on “for free” they would be bumping more profitable workloads off of machines to give away that capacity to people with zero long term loyalty. There is no business reason for google to subsidize these models.
OpenAI has too much money. They’re spending their money in stupid ways.
It makes sense. All these models are money losing businesses.
As a business Google might want to focus on fundamental research 2-3 years from now and not compete on who acquires more money losing customers. Just stay little behind and invest money better.
What you're not seeing are the subsidized Google Cloud startup credits, which includes Gemini. If you're in that program, you choose Gemini because it's essentially "free" and consistent.
I was going to cancel my gemini membership today ..... still going ahead. In my experience, gemini 3.1 pro, 3.5, 3.6 flash constantly lie too much about completing their tasks whereas sol (even though equally dumb) never claims something has been done when it hasn't been.
It's basically a "if we really have to support this for a long time, we want to be compensated for that" pricing strategy. It's about long term maintenance cost being greater _because_ it will be irrelevant.
I want to like Gemini models but my problem thus far has been a lack of coding chops. They still make mistakes, importantly, without correcting them for things like hallucinated API calls or code that doesn't run but they never bothered building or running. I know a lot of this can be fixed with workflows but it still feels like a failing.
GPT-5.6 or Claude models haven't delivered to me non-running code in ages.
Whenever I have Gemini in the flow, it's fast, but mistake riddled. I have low confidence in the output.
I've had some success with Opus driving Gemini models. It's pointless for GPT family since Sol is cheap enough or can drive terra/luna for arguably better performance, same speed, and better outcome.
As for all of the talk in this thread about modalities. Every SOTA model takes screenshots and verifies work now. Grok-4.6 does this, Luna does it, etc. They can also all work _from_ a screen shot or mockup provided.
I don't think it's a major selling point when every model can do it well and reasonably fast.
That said, eagerly awaiting "pro" and improvements to antigravity.
3.7 Flash gets 56 on AA up from 52 for 3.6 Flash. But it seems like this is at the cost of more output tokens per task: 3.6 Flash is 26k, 3.7 Flash is 37k. Due to 3.7 Flash's 2x slashed pricing it's still cheaper per task.
Maybe this is just my experience, but have people had trouble with 3.6 Flash just... getting things it has seen in its context correct? I don't know if it's been insanely benchmaxxed or what, but it'll pull information from websites and immediately get it wrong the token after. Or for example (this is something that happened like yesterday) I asked it to compare the uses of A and B in a language I was learning, and the way I typed it was "Please compare how these two are compared differently: A VS B", and then... it proceeded to compare "VS" and "B". I'm not kidding.
Personally whenever I use Gemini I've just been using 3.1 Pro because I've had insane trouble with them getting things incorrect like this. Hopefully they'll fix it soon / they've fixed it with 3.7 Flash.
I just use web chat as "harness"(lol) or interface and I have mostly switched to Gemini as the free limit basically never run out for me unlike ChatGPT and Claude.
It's on Google AI Studio, which I use for free when I'm not on computers I control.
It did fine on my usual benchmark about configuring old Sparc hardware, maybe output slightly faster than before. Even included something new to check in the firmware.
Has anyone noticed that antigravity has been working really well for the last few weeks. Now with this model it should be working much better. Hope the Google AI Pro Subscription can be used to do some real agentic coding now.
> * For 3.6 and 3.7 Flash, introductory price expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply.
this is hilarious. it is not 2025 any more, by Jan 2027 there will be at least 3 newer generation of models (from other provider) released already. nobody would use flash 3.7 at that time.
sure we used to cling to gemini models in the past, demanding 2.5 models to continue to serve, but since google betrayed us with those price hike, people already spent their time making their production pipeline less dependent on google since then.
heck, even now I'm not sure I even care if they cut the pricing even lower. there are too many models with cheaper price and similar performance now.
They should call it 'face saving pricing after we realized just how terribly did we mis-price the flash 3.5'
> since google betrayed us with those price hike, people already spent their time making their production pipeline less dependent on google since then.
This is my first hand experience. I spent at least $3000 on gemini-3-flash-preview. And exactly $0 total on (3.5+3.6+3.7)
Maybe the business model is to break even on bleeding edge models while making money on the long tail of usage once systems are tuned for a specific model and running in production.
Think of it this way: you are at an enterprise business. You have a workflow implemented a year ago that is working just fine. Swapping out the model for a new one changes behavior in unpredictable ways. Eventually, you'll do it once cost is low enough, but it takes serious labor to validate this, so you'll wait a long enough time for Google to make money.
Prompts can certainly be tuned to a particular model, where updating the model actually results in worse performance. This is perhaps less true today than a year or two ago, but we have seen this on newer models as well. Typically, the less specific the instructions are, the less it's a problem. But sometimes you really need to get into specifics to get good results. Area of work is code porting and translation.
this is becoming less true with every generation of model.
a decent model with a decent harness will determine when the knowledge base is lacking and attempt to fill the holes; thus the good general models can be very easily brought up to speed on niche domains.
Some model are more aggressive by default, some are more verbose by default.
To get the result you want for your specific application, you run experiment with prompts and parameters.
it is partly true, but like I said it is not 2025 anymore. models now get released more often, and still have notable progress so they can safely replace the old models while being faster/cheaper. and thank to chinese models the pricing is pretty much stable and affordable now.
and now we have ai agents to automatic migrate the system with new models. in the past we would need to spend hours to design the prompts, then test the output, then write codes to babysitting it. nowadays any ai agent can do it effortlessly.
I think it's meant to make fun of the fact that Google raised prices on their models and people were upset, and this is Google's way of lowering back the price because by Jan 1st 2027, this model isn't going to be used since people will move on to the latest models.
Personally, I feel like Google blundered on their pricing because while I was using the free version of the Gemini harness, they took away most of the free limits and made people move over to their Anti-Gravity harness for no apparent reason. I was about to splurge for a Pro sub since I already used Google for extra storage but putting up limits like they did made me not want to trust they wouldn't do more price shenanigans. Now their models are behind and it seems like they're scrambling.
Nobody except corporations who built workflows on top of it and don't care about the price because the developer already moved on and nobody wants to touch it.
Does Google believe people want fast models because they have some sort of evidence of that preference? Or are they no longer capable of delivering a Pro model?
I have read that "pro"/"opus"/etc models can actually be worse for everyday coding as they reason "too deeply" and turn over too many stones over-thinking the problem and potentially getting distracted.
This feels absurd to me (my gut is "I want the SMARTEST model I can get!!"), but often I find that my experience of using a flash/sonnet model for every-day workhorse coding they are better.
Its not the same thing, but when I think of that I am reminded of working with some engineers in the past who are incredibly smart and have PhDs (or to put it another way, over-qualified) and they were crap engineers because they'd just not be able to focus on the task and ONLY the task at hand and would get easily distracted by the "why" or "more interesting" things when I just asked them to fix a simple bug or whatever. Again, its not the same thing at all, but it certainly comes to mind when I think of this or experience a pro/opus model suggesting we make huge refactors when a tactical fix is all that is required etc.
Of course, the opus-sized models are great when it comes to huge comprehension/research/debugging efforts where the deeper reasoning is actually useful.
I find Fable completely unbeatable for anything code-related. It's the only frontier model that seems to come with sane defaults.
If it implements something simple like a file export, it just knows that the file should have a meaningful name. Vibecoded feature beats most software's lazy "untitled.png".
So, yes, I want the smartest model even for simple stuff. Maybe especially for simple stuff because the tokens burned will be trivial so the cost doesn't give lower models a comparative advantage.
True, if you have a codebase that works in practice but has dozens of loose ends and poorly defined edge cases than it can chase off into rabbit holes because "oh wait, what if x is undefined instead of null? How is y defined? This outdated package has long known severe security holes and should not be used anymore, do we actually need it?".
Probably both. Having a strong frontier model is necessary not just for the model itself but because it provides a halo effect for your entire line. So if Google could deliver a pro model they would. But I also think Google is targeting the wider market and not picking verticals like Anthropic does. A good enough model is good enough for most generalist tasks, and being fast and cheap is more important to less sophisticated users. Also can't forget Google is at every level of the AI vertical. They're not losing sleep because they're not competitive at the one level in which open weight models come out with the quickness. It reflects poorly on them, and from a marketing perspective its not good but in some ways its actually the least valuable place to be.
If you think about Google and their business/reach, fast and light models suite them the best.
Google probably crunches more tokens daily than the other labs combined, just because basically the entire global population uses Google (sans china) and Google has shoved Gemini into everything.
throughout history, Google has been obsessed with speed as a feature. that was a huge reason people used google search, and then chrome in the first place, and it think its really underestimated by people. Jeff Dean specifcally seems to think about this alot.
That would be concerning if true, since they seem to have made a heavy bet on multi-modal as the way forward.
I wonder if this counts as evidence against that hypothesis? That multi-modal is struggling to keep up with SotA and the best they can offer is competent and fast?
Did the company fix the high friction between any service and their models' API?
I hope so. It seems mind boggling to me that an user needs to surf around different sections (plural) of google cloud console, then this Vertex and do a dozen clicks to issue a simple key.
You can use the Gemini API which is independent of the more complex Vertex AI API. Not sure whether you still have to visit the Google Cloud UI for some things (like billing) though.
At the discounted rates, upgrading from 3 Flash to 3.7 Flash is finally reasonable.
In my evals 3.6 Flash (pre price change) was usually a bit more token efficient than 3 Flash, so I‘m expecting same or even lower cost-per-task on 3.7.
Well, you do get 1 million tokens and the ability to reason over video natively and many of us are forced to pay for 20usd plan anyway due to google drive 5TB, not to mention notebooklm, so it’s not a nothing burguer, it’s just an almost nothing burguer
Offering a 'temporary introductory discount' until Dec 2026 on an LLM is hilarious. In this market, by Jan 2027 this model will be superseded by 5 different providers offering 10x the performance at half the post-discount price anyway.
I wasn't aware of this. Seems Google is lagging the big 3 (Anthropic, xAI, OpenAI) when it comes to frontier models for programming and hard problem solving.
I guess Google's betting on consumers being price-elastic (preferring to tradeoff intelligence for significant cost savings)
After being stuck with using GPT-5.6 models for the past few weeks, I have renewed faith in Google and everyone but OpenAI. The GPT-5.6 models are quite obviously benchmarkmaxxed to make they seem like they are intelligent but they are quite dumb outside anything that not a benchmarked task.
I also think Google is still the best at fitting the most overall intelligences into their models, but for some reason it seems like the model architecture is just bad.
Like, I understand everything, but by this time I don't give anything about any of those announcements.
Theoretically there is some difference between Fable and Opus or Grok and GPT, but at the end of the day I'd look at the bottom left of my screen and to my amusement find out that for the past 3-4 hours I've been using model ______.
If the results are semi-decent, I'd keep it on, if not - I'd randomly switch the model and try again.
Actual thing that would affect my selection would be a number of unused tokens I have left for a model ____ for this week.
Maybe it's cause I'm using those for programming and log parsing and all of them are decent enough, but other than that - there are no leaps I see.
Grok, Meta, Gemini and others all released updates to their models within around a month or two from their respective last release and made significant jumps in benchmarks all around the same time. Any guesses as to why that is? Is it just the release season and/or everyone is benchmaxxing?
its essentially the same model being trained continuously 24/7 with the company periodically publishing just a new checkpoint
each new checkpoint can benefit from better reasoning training, RL on specific tasks and more synthetic data
So why do they seem to release around the same time ?
my guess is because they time major releases around quarterly earnings, investor meetings and other important business milestones.
Once one company announces a major update, the others also have an incentive to ship their latest checkpoint rather than look like they r falling behind.
Sure, they are just checkpoints, that much I guess is obvious. The question is why did they not do frequent releases like this before and why are they making significant jumps in benchmarks so fast and all these companies suddenly falling into that pattern? Earning reports are not to come until end of October, that's not it.
possibly the beginning of the recursive feedback as models begin to aid in their own improvement? especially algorithmic improvements, which seems to have a lot of wide open space for gains
I'm really curious about this: the foundational paper behind today's LLMs came from Google, and some of the world's best scientists were at Google. So why are they falling so far behind in the AI race?
The "let's make money by selling/renting out TPUs" faction has won and the "let's make money by training and selling a frontier model" faction has lost.
And it's arguably not crazy, at least if SemiAnalysis's estimates are to be believed:
* 20% of all TPU shipments from Q3 2026 through Q4 2027 are sold to SPVs serving Anthropic ($150B of contracted revenue); vs
* ~$12B ARR for Gemini.
> The "let's make money by selling/renting out TPUs" faction has won and the "let's make money by training and selling a frontier model" faction has lost.
It's a variation of opportunity cost. A company that has an opportunity to take $1 and make $1.50 on it can't justify an opportunity to spend $1 and make $1.25, even though a less profitable company may make a good living on that. When considering capital allocation, Google has to consider the opportunity cost of investing more into their highly lucrative ads business. Another company that has no access to such a lucrative business uses different opportunity cost when it comes to allocating capital. It can easily be the case that Google could end up justify being in the business of renting out shovels and end up chased out of the business of using the shovels to create AIs entirely because that turns out not to be where the money is. I'm not saying that's obviously inevitable; I'm saying it's a possible and reasonable outcome.
That's why even though the industry produces giants, these giants can never just eat everything. Even though it seems like they have all the money, it isn't practical for them to try to do everything and in fact limits get hit very quickly for anything other than the primary, lucrative business.
Apparently there is no snappy term for this in the business space, according to such AI searches as I have run.
With all due respect, did you read my comment beyond the first paragraph? It addresses both points, TPU economics/pivot to sales + internal shortages making it hard to train models, to the extent they can be addressed based on public sources.
There are other factors at play, but they're more recent/second-order.
Here's a image->html test. Gemini has always swung above its weight class for vision work, so I'm always eager to try it with this.
Original images: https://image.non.io/neonRamenDesigns.webp
Gemini 3.7 build: https://html.non.io/neonRamenGemini3.7
Opus 5 build for comparison: https://html.non.io/neonRamen
Opus is still best in class for this, but it's worth noting how well Gemini 3.7 does vs a more comparable LLM price wise, which is Grok 4.6: https://html.non.io/neonRamenGrok4.6 . I thought Gemini would blow Grok out of the water (it generally has in the past), but Grok has really caught up.
Other thoughts: I really think Google has fallen behind here. Even as a high speed offering (this build took ~7min, which is pretty good!), it wont be able to claim dominance for long with cerebras announcing the Sol preview today: https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultraf... .
It's not a bad model by any means, but I just don't know what situation I'd reach for 3.7 Flash first for. Google really needs a differentiator, especially given how hard it is to get an API key from them. They can't be high friction and non-pareto.
Can you help me understand how it is hard to get an API key from Google? You just head on over to http://aistudio.google.com/api-keys and create a key... not any different from platform.openai.com?
Disclaimer: I work in Google so it might be that this link is not publicly well known
Disclaimer that I haven't tried this since January, so things may have changed in the last 7mo, but this was my experience at that time: https://x.com/pwnies/status/2010523020629274723
At a high level though, as a rule of thumb Google assumes that they're serving companies at Google scale first, and at a human scale second. For other companies it's the opposite. Generally what that means is the first experience you get with a Google product will route you through 8 different dashboards to set up ACLs before you've hired your 2nd employee.
FWIW I definitely did not not need to do anything like that to generate a key via AI Studio. It was like three clicks to get the free tier key, later on enabling billing was a few more plus typing in credit card info.
That is easy but I’ve also found myself in account setup dashboards that were obviously geared toward enterprise trying to set up access to tinker with something AI related. It might have been TTS but it’s been a little while and I can’t quite remember.
it worked for me okay when I needed it for myself in my personal account. But when I tried setup this for a company I spent almost a day solving lot of small puzzles in GCE like how to tell CEO that he have to connect billing account created for other purposes (and he not even remember at time that it exist) to new project and all other things that others talking about.
You need to create a Google Cloud project to create an api key and when you try to create one you very often get error messages like:
“Failed to create project, The request is suspicious. Please try again” or “ You do not have permission to create a key in this project”. You can then navigate multiple screens in GCP to make it work but it’s a hassle compared to any other provider (OAI/Ant/OpenRouter or any of the Chinese labs).
I didn't have that issue back when I originally created my API keys a couple years ago. Just out of curiosity I switched to a different Google account that had never interacted with AI Studio and never used Google Cloud console.
It was literally two clicks, and didn't even leave the page: the dialog asked to create a project and type in a name, I did that, clicked submit and then it was selected as the default project. One more click and I had the free tier API key.
Not saying you didn't have that experience at the time, but personally I have had zero issues with AI Studio and consider it the most simple/fastest dev dashboard to get started compared to the others like OpenAI/Anthropic (thanks to Google's free tier that lets you skip billing setup annoyances just to play around with Gemini).
Despite what HN threads (that are also frequently confused and talking about GCP instead) portray as universal/widespread issues or the process being complex and time consuming somehow.
GCP/vertex is a maze
I haven’t tried in about a year, but I could never do any meaningful work outside 1P Google apps (antigravity) due to such fast throttling.
On top of being the hardest website to navigate, Google console a) doesn't have real time billing (!) b) doesn't allow you to set a budget limit.
Sorry but it's not worth waking up with a 100k$ bill, fix your platform first.
You prepay for tokens exactly like the OpenAI/Anthropic dev dashboards when using AI Studio, which the link above is pointing to not GCP.
https://ai.google.dev/gemini-api/docs/billing#spend-caps
I use LLMs rarely, and only for digging into subjects which I can't find enough information using search engines. I only tried Claude and Gemini, but Gemini both returns faster and higher quality information which I can use for more targeted digging myself.
Google being Google, their models tend to be better at finding, organizing and presenting information, from my experience.
Sol on Cerebras is going to be expensive AF
Moving from either frontier intelligence or frontier latency to a single model that does both at the same time is potentially a game changer in certain industries. I can easily see e.g. hedge funds dropping tons of money on this, because it means they can now do the same thing as their competitors, but much faster. That's basically a license to print money.
Depends on the definition of friction. If someone is in the Google ecosystem, why would they reach out of it.
I already use GCP and Google for work, and getting an API key was so annoying that even I couldn't be bothered after a while of looking around.
Maybe things there have improved some, but when I was looking it was a huge runaround.
Hmm. My company has an internal portal for generating Gemini API keys. I select a project from a drop down, enter a name, and press okay.
I like using 3.5-flash-lite for doing cheap PDF and Image data extraction stuff. I don't think there is a better bang / buck model right now (3.1 is cheaper but a lot worse).
5.6 Luna costs far less and benchmarks far better, have you compared for this task?
Probably cheaper to run a Mac Mini with VisionKit (private APIs if you need bounding rects).
I think both outputs are really good. I don't see a lot of differences. So what exactly should be looking at and notice that one model did worse or better than the other one.
EDIT: OKAY I see it's mostly the "image" generation, not so much the HTML... Noticeable in the food photos and the foodtruck/cart photo
How are you doing this with Opus. Clearly I’m missing something. I always turn to ChatGPT when I need images because Opus typically refuses. I’ve tried Claude Code and Claude online in the past. I’m pretty sure neither created images for me and I thought this was because Anthropic was focused on code.
I guess I need to try harder. :)
Images and build step were generated with my own tool (https://news.ycombinator.com/item?id=48995754 - it's why I'm often running these img->html tests).
Opus can't generate images since A\ doesn't have a diffusion model.
did you build your own diffusion model?
I have a few custom ones (a post-trained flux 2 checkpoint for web design and a image->metalness map generator that the build step can call for more advanced lighting situations), but gpt-image-2 is better than my own for design, so it's weighted much more heavily in outputs my tool generates. I think gpt-image-2 currently generates 99%+ of the design outputs on diffui
I believe they are testing giving it an image, which you can do in Claude code by dragging/dropping into the terminal or copy/pasting, and asking it to build the html equivalent.
There is some irony being a developer and reading along the lines of: "oh look at the comparison between these models executing a task for a few cents on a job i'd be charging 1k minimum"
FYI, developers are rarely given such a rich UX mock.
Depends who you work with, what's the intention, budget, etc. I'd agree this is a really good one.
I'm used to incremental Figma wireframe -> final product and working together with a designer.
I don't think I'd say rarely. Companies rarely allocate the design resources to produce that, but the companies that do are typically much larger, so the actual number of individual developers that get rich mocks is probably closer to 40-50%.
I'm curious how much the harness plays into this. I'm somewhat surprised by the gemini and grok results, they seem to have strongly deviated from the original images. I'm thinking maybe the harness has a big effect? It's possible to proxy in different models to claude code, if you're curious you might find it interesting to test!
Harness could be a part of it, but worth noting both the Opus and Gemini 3.7 flash tests were both ran through opencode.
The grok test was ran through the cursor cli agent however.
why Grok not through `Grok Build` ?
I'm not sure what prompt you put in but did Gemini replace the all of the images in the original with its own? That would be really weird behavior unprompted.
The prompt is a build step generated by my tool for image->html conversion, which includes APIs the model can call to generate images/patterns/svgs.
https://image.non.io/12275ee8-71e9-4941-823b-e51fec157b4d.we...
The agent is told to generate assets as part of the buildout. It gets to decide what the prompt is for them / whether to do postprocessing like background removal / what type of asset to generate.
They both have horizontal scroll on mobile...
Oh yea, as a disclaimer the models didn't have any instructions to do a mobile version. I haven't tested them on mobile at all.
The "introductory pricing" for this 3.7 Flash model is really weird.
It's scheduled to double in price on December 31, 2026, but who would anticipate still using this model five months from now? Especially since 3.6 Flash came out just three weeks ago!
My first effort with default thinking level produced an ambitious pelican, let down by a flawed bicycle: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
Then I ran it on high, medium and low thinking levels (oddly minimal is no longer an option, which WAS an option for 3.5 and 3.6) and got a pretty excellent pelican for the first two:
https://tools.simonwillison.net/markdown-svg-renderer.html#u...
UPDATE: That was in Safari, but as pointed out in the replies here the pelicans do NOT render well in Firefox or Chrome! Best guess is that's because of this invalid filter in the SVG:
Filters are meant to contain additional elements, not be empty: https://drafts.csswg.org/filter-effects/#FilterElement - so maybe Chrome and Firefox remove the element that references the broken filter but Safari doesn't?At $work we still have some places using Opus 4.5. Even places using Qwen 2.5-VL, which is now 18 months old. It works, and upgrading is work (we'd have to validate the new model performs comparable in all the corner cases that currently work just fine)
Those kind of workloads would be hit by an end of introductory pricing. And it's exactly the kind of cases that are not very price sensitive. Where we are price sensitive we track new model releases closely, where we aren't other issues get priority as long as llm performance is good enough
> got a pretty excellent pelican for the first two
This suggests you primarily use Safari.
While the bike renders, the pelican doesn’t in Chrome and Firefox.
Probably one of the more serious defects I’ve seen with the pelican. It’s one thing when animated SVGs have bugs, but another when plain ones do.
my whole world is shifting. have I been seeing _different pelicans_ from everyone else?!
... whoa, that's true! Thanks for catching that.
I think introductory here means more or less permanent but they can't publicly admit there are no takers at a higher price.
Anthropic for instance announced a couple of days ago that they are making Sonnet's 'introductory pricing' permanent https://xcancel.com/claudeai/status/2086891169217122586
maybe exactly because the frontier moves, introductory pricing makes sense as you want to free up compute for the newer models
This pelican gave me a good laugh, because there's enough reasoning that the render is out of sight initially. The buildup!
It's not weird if you're in marketing.
Ever since the insane discount with GPT-5.6 Luna, not much excites me anymore. I mean just look at the benchmarks, even though Gemini 3.7 Flash performs well on the DeepSWE 1.1, Luna (Max) still performs way better. I personally have stuck to Luna (Xhigh) because its been more than enough and does not bloat up the context window too fast with reasoning tokens.
https://deepswe.datacurve.ai
> Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply.
Compare this to Luna which is at $0.2/1M input ($0.02 cached) and $1.2/1M output.
https://developers.openai.com/api/docs/models/gpt-5.6-luna
GPT-5.6 Luna is an insanely powerful model for its price. It's been great for coding workflows where I guide the LLM's hand step by step. It's also insane to see my weekly limit drop by than 2% after an hour of coding ever since the discount.
However, I've noticed 2 drawbacks with Luna. Context rot is much more palpable than Terra and Sol. It tends to get confused and go into rabbit holes when it's context gets filled up. In addition, when instructions are vague, it performs poorly and tends to write way to more code than necessary, but that is to be expected of smaller models. In all, for clearly defined, bite-sized coding tasks, Luna's price-to-performance has been insane. It might have very well commanded the price tag of Sol if it came out just a year ago.
luna is the first model that has outdone gpt-5-mini on the pareto frontier for some of my high value, cost sensitive ai product workflows. it's both cheaper (by about 60% in real world use) and higher quality based on my test harnesses. I was really worried that costs would go up since there wasn't a replacement as of a few weeks ago and gpt-5-mini is scheduled to be sunset toward the end of the year. So long as they don't randomly sunset this model anytime soon, that worry has now subsided.
I practically switched to doing everything with Luna or DeepSeek V4 flash. I haven't feel the need for the more expensive models.
I am in the same boat as you. I am using Luna and DeepSeek Flash. Both super fast, super cheap, and I have not felt need for anything more capable in few weeks.
I'm curious how much people are manually curating context these days; I'm increasingly feeling for myself that it being auto-managed inside a front-end like claude code is not ideal, and I'd rather have more control over what exact files and pieces of discovery go into a particular prompt, and the ability to more easily "fork" a session and ask asides or make notes/todos in a way that doesn't disrupt or confuse a more focused task going on.
I don't think I want a gastown-style "just yolo everything" approach, in fact I really want more control over how decisions are made and with what info. Does this exist?
I am not sure this enlightens you with anything but I have a TODO.md file with three headlines. Todo, Doing and Done. The agent is aware of it and knows on which task we are on.
On complete, it moves the user story from Doing to Done. I also have a MEMORY.md file that the agent read and writes in the beginning of a new conversation and at the end of our conversation to update stale information. These files are referred to every time I start a new conversation.
Regarding forking, I know Codex has such button underneath each message that lets you fork the whole conversation. I usually do that when I want to sidetrack and discuss something.
I use no SKILLS or commands like /goal. I’ve come a long way with just prompts and markdown files. Its all a different way of encapsulating instructions anyways.
> does not bloat up the context window too fast with reasoning tokens
How much does that matter if it's reset at every turn?
what do you mean by reset at every turn? context stays until compaction. if you remove the reasoning tokens after every turn you will be constantly blowing cache which is far worse than filling up context.
Does it reset at every turn? From my experience in Codex for example, Luna (Max) fills the 256k token window relatively quick. The only thing lowering the context window again is the compaction.
Benchmarks mean very little. The difference between Luna and Sol in the real world is massive.
It’s a big difference but Luna is very usable. I’ve plugged it into the slot I used to have GLM 5.2 in; I think it’s just as good. And it is less costly. I have Sol do planning and design but do most task execution with Luna now.
They need to release benchmarks against Luna/Terra. Luna is much cheaper which feels like it undercuts the need for Flash.
I've always considered the Flash series of models to be for low-cost, high-volume, mostly text-based use cases (e.g. summarization, parsing, formatting), emphasis on low-cost.
[edit: ah, benchmarks here: https://blog.google/innovation-and-ai/models-and-research/ge...
more of a Terra than Luna competitor which is an interesting positioning. I feel like differentiation at the mid-tier of models is pretty difficult.]
flash-lite is more of their luna tier competitor but even still not quite there yet, but gemini's dominance on multimodal and image understanding i think really gets downplayed on this site when most people think the only think you can do with LLMs is write code
If anyone knows of a cheaper vision llm with the same accuracy I would love to switch
That's been my association as well. I see Flash get brought up a lot in relation to things like OCR and PDF processing frequently, and a lot of other routine multimodal workloads.
Yes this is my impression as well. To be fair I didn't compare to Luna yet, but Gemini 3.5 Lite is a very good and cheap multi-modal data extraction model.
gemini flash is probably the best model for visual tasks right now. they also make it really easy to ingest videos
Also best at OpenSCAD, seemingly for the same reason, at least in terms of "iterate on a design, comparing visual output to target".
Crazy it's still the only video understanding endpoint. It's what I use it for and no other model even offers a competitor.
Yes, was going to say I use it exclusively for video and audio. The ability to give it a YouTube link through the API and ask questions about it is awesome
Ah, multimodal is a great point. I'll need to try that some time.
They compared against 5.6-terra on the model card: https://deepmind.google/models/model-cards/gemini-3-7-flash/
Matched roughly with Sol on DeepSwe cost per task.
Luna way cheaper. DeepSeek used to be, but I think it's somewhere on Sol's curve after the price hike.
On DeepSwe it's strictly beaten by Luna on max, cost and result.
Damn, Luna on max is as good on DeepSWE as Kimi k3, I think I dismissed this model unjustly.
Why is Gemini represented by points on this cost-quality plane, while competitor's models are represented by curves?
Competitors release multpile models and their curves reflect reasoning effort of each single model.
Gemini doesn't have adjustable reasoning effort (at least on the graph) so each of its curves is just one point.
Gemini has levels though
Actual announcement: https://blog.google/innovation-and-ai/models-and-research/ge...
So it's better than 3.6 Flash, at half the price. I've been pretty excited about Gemini models recently, they just feel so fast after spending most of the day at work waiting for Opus 5.
Yes, 3.6 Flash is very fast. I used to get a fair amount of usage of the Gemini Flash models on the free tier. I signed up for their $4.99/month tier (includes 400GB of Google space which was also enticing) and it turns out I only get about 15 to 20 minutes of usage before I get a come-back-in-7-days message. Comically low usage limits on that plan.
> So it's better than 3.6 Flash, at half the price.
I think its the same price..
They seem to have reduced the price of 3.6 Flash at the same time that they released this 3.7 model:
https://ai.google.dev/gemini-api/docs/pricing today has 3.6 Flash at $0.75/$3.75 until December 31st 2026, then doubling.
https://web.archive.org/web/20260809105129/https://ai.google... Internet Archive copy of that page from 9th August has 3.6 listed at $1.50/$7.50 with no mention of the price changing.
The Gemini Flash models makes perfect sense to me coming from a company like Google. Google AI Mode for search is a product I really find useful. It makes sense that Google focused on smaller, faster yet smart enough models that wouldn't break your bank on inference. It plays well into their product ecosystem.
Google AI Mode consistently gets me consistently good results and good speeds. It really changes what "googling" is for me.
I like Gemini (I'm just a dumb person without knowledge of 'benchmarks' or how x compares to y) while understanding that shoving it into Google auto summaries has been a bad idea and produces inaccurate results
https://artificialanalysis.ai/models/gemini-3-7-flash
The selling point for gemini continues to be speed and particularly end-to-end response time.
It's funny that they don't mention this at all in the marketing or tech specs when it's obviously the biggest selling point by far. Without this it would be completely irrelevant.
Worth noting that OpenAI just announced that they got the full GPT 5.6 Sol model running on Cerebras at 750 tokens per second. No announcement of the pricing though...
> It's funny that they don't mention this at all in the marketing or tech specs when it's obviously the biggest selling point by far. Without this it would be completely irrelevant.
Good catch! You're right to point that out. My previous marketing copy missed that specific detail. Thank you for bringing it up!
Cerebras is crazy to watch on GPT OSS or Gemma, I feel like we need a new VibeOS demo but with Cerebras, the OS would literally build itself in a few seconds.
https://youtu.be/7NfyZhV1dKM?t=52
I use this in a customer facing application and Gemini’s speed makes the experience feel much better.
The application isn’t so complicated that you need opus level reasoning or code writing, we need “good enough” data retrieval and processing with natural language queries and the ability to answer follow up questions.
For that Gemini works well for a decent price.
It's like the Intel Optane of AI
I've blown away by flash 3.6's speed while Opus chugs along for _hours_ on similar tasks. I've gotten into a opus designed -> gemini implemented -> opus reviewed dev cycle recently.
I am actively using Gemini flash to "translate" what Opus says into human language. I let opus do the design (with my assistance) and implementation, but then the report that Opus writes gets translated by Gemini so that I don't have to waste time to understand it.
That’s like the army guy in movies from the 90s who shouts “IN ENGLISH, PLEASE!” after the scientist explains the conflict of the plot.
What's the typical response time for Gemini compared to other models?
On my benchmark where AIs generate ~20 different 3D models about 1/2 the time of Opus and 1/3 of the time of Kimi K3 and 2/3 of time of sonnet.
Sol high is almost the same speed if you take into account drastically lower token use. Look at the artificial analysis speed vs token use. Gemini is 7x faster but 5x more tokens. And that's with Sol high being a substantially better model.
Edit: and Sol medium actually has the same AA intelligence score as Gemini 3.7, and has >7x fewer tokens, actually making it faster
is presumes you are doing longer difficult agentic tasks, if youre doing a simple problem in 1 or 2 shots, not really multi turn then theres no comparison.
The multimodal abilities are great, but if you deal with text only, what is the benefit of using this over DS V4 Flash/Pro? 13-26x cheaper with comparable intelligence, and available across many different inference providers.
I fail to see the usecase where DS V4 Pro is not enough, but Flash 3.7 is - except multimodal.
Luna is similar, and also 8x cheaper. Source: artificialanalysis
The only benefit I can see is the speed, that looks to be outstanding, probably thanks to their TPUs.
for non-coding applications, i think speed is a real differentiator. Im building an app that uses LLMs for some functionality that the user would not have any reason to expect is using AI and therefore having then wait seconds or minutes is just not feasible. latency is a huge upside for me
Anything interacting with the real world seems like latency would be hugely important. Something more asynchronous friendly (like coding) is for obvious reasons over represented here
> 13-26x cheaper with comparable intelligence, and available across many different inference providers.
Well, compared to 2 months ago, it's no longer 100x more expensive for similar levels of quality...
If they continue monthly-ish releases by 3.9 - by Halloween - they should be close to the best in terms of what you get for what you pay for.
In 2 months, they've gone from basically the bottom of the pack to at least being somewhat usable and competitive.
OpenAI and Anthropic release in a month, and change things. OpenAI is claiming to be close to an Astra release - but that seems like a Fable type release - where they're just releasing a better more expensive model, not more cost effective models.
From my own testing, Gemini 3.5/3.6 Flash is better than DS v4 Flash/Pro on text ability.
DSV4 Flash is in a tier of its own, until at least the price change arrives.
did you try to ingest 1M documents per hour with any provider except GCP with Flash? None work at scale. Deepseek, Luna, Mistral all fail. 1 in 3 requests is a fail. I stopped trying.
The only thing that works at scale is gemini flash.
I guess the question then becomes "are you sure you'll do text only?"
I could probably do text only for my workflow (feature development/debugging for web microservices) but sometimes it is easier to just toss a screenshot into the Claude prompt, so that gives it an edge.
If your workflow is 100%, certifiably never ever going to involve an image, then yeah, this isn't going to be huge.
> over DS V4 Flash/Pro? 13-26x cheaper with comparable intelligence, and available across many different inference providers.
That's why DS4 already had a huge price hike announcement.
The inference providers did not raise the prices no?
Deepseek as a company can just increase prices for the crazily cheap cache they have, that's their only lever.
I guess the demand is just too high... But even after the price hike, ds is still much cheaper?
Gemini Flash is one of the best "good-enough" models. I use this type of model daily, for automation and quick development iteration loops.
Unfortunately, it's often not strong enough for heavy refactoring and long running development loops.
Yeah we use it for auto-triage of incidents, attempts to auto-remediate, and escalation to human. But for actual development, it’s not a viable option for us.
'Tis a good workhouse, indeed. I hope they give us a 4.0 Pro that can use Flash subagents soon.
Yup. I use it for a ton of mundane queries (stuff that I might have used Google search for in the past) and it's great. Nice and fast and correct more often than not, especially if you prompt it in a way that it invokes Google search (but filters out ads and SEO slop). It's even alright at programming tasks but if it stumbles then I'll escalate to Gemini Pro with extended thinking.
> What's new in Gemini 3.7 Flash [0]
> Coding and agentic tasks: Significantly higher quality on real-world software engineering and agentic benchmarks, improving issue resolution and reducing failed agent loops.
> Web development and stronger design parity: Generates higher-fidelity desktop and web application code directly from design mocks, with strong gains in design adherence and in auditing existing codebases against mocks to verify 1:1 design parity.
> Promotional pricing: Gemini 3.7 Flash will be available at an introductory price of $0.75/1M input tokens and $3.75/1M output tokens. We’re also applying this new rate to 3.6 Flash. Introductory pricing expires on December 31, 2026; after, $1.50/1M input tokens and $7.50/1M output tokens will apply.
Still no sign of 3.5 Pro. Will have to test it, low expectations given every other model from the Gemini 3 lineage, but one can hope. Just struggle to understand the promotional pricing being temporary for four months. Given this industry, I'd be hard pressed if 3.7 Flash was still in use by end of year, so why not make it the official pricing?
[0] https://ai.google.dev/gemini-api/docs/latest-model
>Given this industry, I'd be hard pressed if 3.7 Flash was still in use by end of year, so why not make it the official pricing
It was probably to placate some kind of general internal pricing/revenue benchmark that doesn't account for new model releases. Politicians do shit like this incessantly and it reeks of bureaucracy.
I suspect it's a bit of a signal to investors etc.
"Hey, we are not in a race to the bottom. This is our usual pricing, but this now is a promotion because we know we're coming from behind and need to entice users."
They're drawing a line in the sand on monetisation and signalling that to everyone, while in reality offering it a deep discount (no idea if profitable or not) knowing that this model will probably be obsolete before then.
So they keep pushing these Flash models because they don’t really have a powerful model…or better said their ‘pro’ model is actually a flash
I don't get it, Google could heavily subsidy their Gemini models to make it more attractive, but they prefer to not do it. I don't know one soul who is using Gemini models to code. Even OpenAI who doesn't have money or capacity is offering their Luna model at $1.2 per 1M/out.
Why would they? Unless they have lots of unused tpu real estate that they could host it on “for free” they would be bumping more profitable workloads off of machines to give away that capacity to people with zero long term loyalty. There is no business reason for google to subsidize these models.
OpenAI has too much money. They’re spending their money in stupid ways.
It makes sense. All these models are money losing businesses.
As a business Google might want to focus on fundamental research 2-3 years from now and not compete on who acquires more money losing customers. Just stay little behind and invest money better.
What you're not seeing are the subsidized Google Cloud startup credits, which includes Gemini. If you're in that program, you choose Gemini because it's essentially "free" and consistent.
I was going to cancel my gemini membership today ..... still going ahead. In my experience, gemini 3.1 pro, 3.5, 3.6 flash constantly lie too much about completing their tasks whereas sol (even though equally dumb) never claims something has been done when it hasn't been.
Two thoughts:
1) It makes sense to try 3.7 flash before cancelling.
2) Prompting models to be honest is surprisingly effective in my recent experience. But only if they listen to instructions.
> 3.7 Flash is available through the end of the year at an introductory price of $0.75/1M input tokens and $3.75/1M output tokens.
> Introductory pricing expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply.
Clearly this model will be irrelevant by Jan. 2027, why would Google even bother to say this?
It's basically a "if we really have to support this for a long time, we want to be compensated for that" pricing strategy. It's about long term maintenance cost being greater _because_ it will be irrelevant.
They want to maintain the perception that Flash is worth $7.5/mot, so they can charge more for the next one.
Its probably just a corporate symptom, weird stuff like this happens in messy large orgs.
Maybe they know something we don't. What if all frontier lab do this? Maybe this is actual cost of running these llm.
I want to like Gemini models but my problem thus far has been a lack of coding chops. They still make mistakes, importantly, without correcting them for things like hallucinated API calls or code that doesn't run but they never bothered building or running. I know a lot of this can be fixed with workflows but it still feels like a failing.
GPT-5.6 or Claude models haven't delivered to me non-running code in ages.
Whenever I have Gemini in the flow, it's fast, but mistake riddled. I have low confidence in the output.
I've had some success with Opus driving Gemini models. It's pointless for GPT family since Sol is cheap enough or can drive terra/luna for arguably better performance, same speed, and better outcome.
As for all of the talk in this thread about modalities. Every SOTA model takes screenshots and verifies work now. Grok-4.6 does this, Luna does it, etc. They can also all work _from_ a screen shot or mockup provided.
I don't think it's a major selling point when every model can do it well and reasonably fast.
That said, eagerly awaiting "pro" and improvements to antigravity.
Try this one, it's a step jump in coding capabilities for me over 3.6.
3.7 Flash gets 56 on AA up from 52 for 3.6 Flash. But it seems like this is at the cost of more output tokens per task: 3.6 Flash is 26k, 3.7 Flash is 37k. Due to 3.7 Flash's 2x slashed pricing it's still cheaper per task.
Always loved the Gemini. Helped my work alot
Maybe this is just my experience, but have people had trouble with 3.6 Flash just... getting things it has seen in its context correct? I don't know if it's been insanely benchmaxxed or what, but it'll pull information from websites and immediately get it wrong the token after. Or for example (this is something that happened like yesterday) I asked it to compare the uses of A and B in a language I was learning, and the way I typed it was "Please compare how these two are compared differently: A VS B", and then... it proceeded to compare "VS" and "B". I'm not kidding.
Personally whenever I use Gemini I've just been using 3.1 Pro because I've had insane trouble with them getting things incorrect like this. Hopefully they'll fix it soon / they've fixed it with 3.7 Flash.
I just use web chat as "harness"(lol) or interface and I have mostly switched to Gemini as the free limit basically never run out for me unlike ChatGPT and Claude.
Also impressed with Grok for some stuff.
Introductory price seems a bit weird as it expires at the end of the year and by then it’s going to be significantly outdated.
It's on Google AI Studio, which I use for free when I'm not on computers I control.
It did fine on my usual benchmark about configuring old Sparc hardware, maybe output slightly faster than before. Even included something new to check in the firmware.
For those looking for the full benchmark figures and technical overview, Google's primary announcement post is here: https://blog.google/innovation-and-ai/models-and-research/ge...
Has anyone noticed that antigravity has been working really well for the last few weeks. Now with this model it should be working much better. Hope the Google AI Pro Subscription can be used to do some real agentic coding now.
I have been using the first party version all year, and with this model in it I am really happy. It’s fast and works.
> * For 3.6 and 3.7 Flash, introductory price expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply.
this is hilarious. it is not 2025 any more, by Jan 2027 there will be at least 3 newer generation of models (from other provider) released already. nobody would use flash 3.7 at that time.
sure we used to cling to gemini models in the past, demanding 2.5 models to continue to serve, but since google betrayed us with those price hike, people already spent their time making their production pipeline less dependent on google since then.
heck, even now I'm not sure I even care if they cut the pricing even lower. there are too many models with cheaper price and similar performance now.
> introductory price
They should call it 'face saving pricing after we realized just how terribly did we mis-price the flash 3.5'
> since google betrayed us with those price hike, people already spent their time making their production pipeline less dependent on google since then.
This is my first hand experience. I spent at least $3000 on gemini-3-flash-preview. And exactly $0 total on (3.5+3.6+3.7)
gemini-3-flash-preview is legit amazing and cheap. That's why i spent over 10k on it.
Maybe the business model is to break even on bleeding edge models while making money on the long tail of usage once systems are tuned for a specific model and running in production.
How can system be tuned for a specific model? Model is fungible, often one model strictly greater on both quality and price.
Think of it this way: you are at an enterprise business. You have a workflow implemented a year ago that is working just fine. Swapping out the model for a new one changes behavior in unpredictable ways. Eventually, you'll do it once cost is low enough, but it takes serious labor to validate this, so you'll wait a long enough time for Google to make money.
Prompts can certainly be tuned to a particular model, where updating the model actually results in worse performance. This is perhaps less true today than a year or two ago, but we have seen this on newer models as well. Typically, the less specific the instructions are, the less it's a problem. But sometimes you really need to get into specifics to get good results. Area of work is code porting and translation.
Models are not fungible, if you're building certain types of products on them.
this is becoming less true with every generation of model.
a decent model with a decent harness will determine when the knowledge base is lacking and attempt to fill the holes; thus the good general models can be very easily brought up to speed on niche domains.
Ugh? It is not just the knowledge
Some model are more aggressive by default, some are more verbose by default. To get the result you want for your specific application, you run experiment with prompts and parameters.
You're thinking like an engineer.
Think like a regulator.
The business model is to replace the entire human economy.
it is partly true, but like I said it is not 2025 anymore. models now get released more often, and still have notable progress so they can safely replace the old models while being faster/cheaper. and thank to chinese models the pricing is pretty much stable and affordable now.
and now we have ai agents to automatic migrate the system with new models. in the past we would need to spend hours to design the prompts, then test the output, then write codes to babysitting it. nowadays any ai agent can do it effortlessly.
This is added specifically so you migrate out of those as fast as next models will be available.
I think it's just to signal that prices will go up in the future.
Isn't it a good thing to know about price hikes in advance? If I were building a product around it, I would certainly care.
I think it's meant to make fun of the fact that Google raised prices on their models and people were upset, and this is Google's way of lowering back the price because by Jan 1st 2027, this model isn't going to be used since people will move on to the latest models.
Personally, I feel like Google blundered on their pricing because while I was using the free version of the Gemini harness, they took away most of the free limits and made people move over to their Anti-Gravity harness for no apparent reason. I was about to splurge for a Pro sub since I already used Google for extra storage but putting up limits like they did made me not want to trust they wouldn't do more price shenanigans. Now their models are behind and it seems like they're scrambling.
Nobody except corporations who built workflows on top of it and don't care about the price because the developer already moved on and nobody wants to touch it.
I think this is a play to get around EU regulation about false sales.
What? Jan 2027 is just about four months away. People surely still use models from four months ago today.
Is this the reason why Jeff Dean, Sanjay Ghemawat and other DeepMing, Gemini people got kicked out of Google?
If so, now I understand why they didn't want to release this model
Reposting my comment from the other thread https://news.ycombinator.com/item?id=49288847
They compare it to 5.6 Terra, however https://cognition.com/frontiercode puts Terra at about 1/2 the price
Also have to compare to the recent Grok 4.6 release, which appears to straight up be better AND cheaper
Hard to understand why anyone would choose 3.7 Flash under these conditions.. is Deepmind still a frontier lab?
Gemini Flash 3.6 High was about 10x faster than Luna xhigh for the work that I tested it for, and it got similar results.
[dupe] https://news.ycombinator.com/item?id=49288847 (35 points, 8 comments)
Does Google believe people want fast models because they have some sort of evidence of that preference? Or are they no longer capable of delivering a Pro model?
I have read that "pro"/"opus"/etc models can actually be worse for everyday coding as they reason "too deeply" and turn over too many stones over-thinking the problem and potentially getting distracted.
This feels absurd to me (my gut is "I want the SMARTEST model I can get!!"), but often I find that my experience of using a flash/sonnet model for every-day workhorse coding they are better.
Its not the same thing, but when I think of that I am reminded of working with some engineers in the past who are incredibly smart and have PhDs (or to put it another way, over-qualified) and they were crap engineers because they'd just not be able to focus on the task and ONLY the task at hand and would get easily distracted by the "why" or "more interesting" things when I just asked them to fix a simple bug or whatever. Again, its not the same thing at all, but it certainly comes to mind when I think of this or experience a pro/opus model suggesting we make huge refactors when a tactical fix is all that is required etc.
Of course, the opus-sized models are great when it comes to huge comprehension/research/debugging efforts where the deeper reasoning is actually useful.
I find Fable completely unbeatable for anything code-related. It's the only frontier model that seems to come with sane defaults.
If it implements something simple like a file export, it just knows that the file should have a meaningful name. Vibecoded feature beats most software's lazy "untitled.png".
So, yes, I want the smartest model even for simple stuff. Maybe especially for simple stuff because the tokens burned will be trivial so the cost doesn't give lower models a comparative advantage.
That does not match my own experience, which is why I wonder if Google has evidence of that.
Consistently, lower intelligence models provide worse results in my own work. But I don't have evals on my side, just vibes.
This is my experience at least.
True, if you have a codebase that works in practice but has dozens of loose ends and poorly defined edge cases than it can chase off into rabbit holes because "oh wait, what if x is undefined instead of null? How is y defined? This outdated package has long known severe security holes and should not be used anymore, do we actually need it?".
Probably both. Having a strong frontier model is necessary not just for the model itself but because it provides a halo effect for your entire line. So if Google could deliver a pro model they would. But I also think Google is targeting the wider market and not picking verticals like Anthropic does. A good enough model is good enough for most generalist tasks, and being fast and cheap is more important to less sophisticated users. Also can't forget Google is at every level of the AI vertical. They're not losing sleep because they're not competitive at the one level in which open weight models come out with the quickness. It reflects poorly on them, and from a marketing perspective its not good but in some ways its actually the least valuable place to be.
If you think about Google and their business/reach, fast and light models suite them the best.
Google probably crunches more tokens daily than the other labs combined, just because basically the entire global population uses Google (sans china) and Google has shoved Gemini into everything.
throughout history, Google has been obsessed with speed as a feature. that was a huge reason people used google search, and then chrome in the first place, and it think its really underestimated by people. Jeff Dean specifcally seems to think about this alot.
All the leaks say their latest attempt at a Pro model was not competitive.
That would be concerning if true, since they seem to have made a heavy bet on multi-modal as the way forward.
I wonder if this counts as evidence against that hypothesis? That multi-modal is struggling to keep up with SotA and the best they can offer is competent and fast?
Especially not competitive at software engineering.
what a week - lets see it draw a weird animal doing a weird thing on a bicycle
Shouldn't it be drawing the whole Silmarillion now?
no that's in Flash 3.8 Pro
When are we getting another pro model from Gemini? Or are they simply focusing on the niche of fast but moderately capable models?
So their "Flash" model won't be cheap. Are they gonna make a new one that's cheaper? Gemini-3.8-Silverlight? ;)
Did the company fix the high friction between any service and their models' API?
I hope so. It seems mind boggling to me that an user needs to surf around different sections (plural) of google cloud console, then this Vertex and do a dozen clicks to issue a simple key.
You can use the Gemini API which is independent of the more complex Vertex AI API. Not sure whether you still have to visit the Google Cloud UI for some things (like billing) though.
Model card: https://deepmind.google/models/model-cards/gemini-3-7-flash/
Somewhere in the same neighborhood as GPT 5.6 Tera and Sonnet 5, depending on the bench.
At the discounted rates, upgrading from 3 Flash to 3.7 Flash is finally reasonable.
In my evals 3.6 Flash (pre price change) was usually a bit more token efficient than 3 Flash, so I‘m expecting same or even lower cost-per-task on 3.7.
Maybe a play by Google to deprecate 3 Flash soon.
Well, you do get 1 million tokens and the ability to reason over video natively and many of us are forced to pay for 20usd plan anyway due to google drive 5TB, not to mention notebooklm, so it’s not a nothing burguer, it’s just an almost nothing burguer
Offering a 'temporary introductory discount' until Dec 2026 on an LLM is hilarious. In this market, by Jan 2027 this model will be superseded by 5 different providers offering 10x the performance at half the post-discount price anyway.
How does it compare to Opus 5.0 and Fable 5 for coding? E.g. in Cursor or OpenCode?
It is not a competitor to those it competes with Sonnet.
Google's Opus competitor is 3.1 Pro Preview which is essentially obsolete (competed with Opus 4.6). They do not have a Fable/Sol competitor.
I wasn't aware of this. Seems Google is lagging the big 3 (Anthropic, xAI, OpenAI) when it comes to frontier models for programming and hard problem solving.
I guess Google's betting on consumers being price-elastic (preferring to tradeoff intelligence for significant cost savings)
Thats a unique definition of Big 3
I would replace xAI with Moonshot AI since Kimi K3
After being stuck with using GPT-5.6 models for the past few weeks, I have renewed faith in Google and everyone but OpenAI. The GPT-5.6 models are quite obviously benchmarkmaxxed to make they seem like they are intelligent but they are quite dumb outside anything that not a benchmarked task.
I also think Google is still the best at fitting the most overall intelligences into their models, but for some reason it seems like the model architecture is just bad.
Like, I understand everything, but by this time I don't give anything about any of those announcements.
Theoretically there is some difference between Fable and Opus or Grok and GPT, but at the end of the day I'd look at the bottom left of my screen and to my amusement find out that for the past 3-4 hours I've been using model ______.
If the results are semi-decent, I'd keep it on, if not - I'd randomly switch the model and try again.
Actual thing that would affect my selection would be a number of unused tokens I have left for a model ____ for this week.
Maybe it's cause I'm using those for programming and log parsing and all of them are decent enough, but other than that - there are no leaps I see.
Is that he model that supposed to be Pro, but then they changed their mind?
No, relabelling a Pro model as Flash would make no economic sense (the Pro series is larger than Flash and more expensive to serve).
So, again with a Flash model. Why are they so afraid to put out an actual SOTA frontier high intelligence model?
We still don't have a 3.5 Pro, and along comes 3.7 Flash?!
In August of 2026, Gemini became self-aware, and began producing increasingly crappy flash versions of itself...
Glad to see that the company with the most data is releasing the most amount of models. Some things do make sense
I'm only interested in the state-of-the-art model by each provider.
For Google, this is still gemini-3.1-pro-preview, right?
This is all a naming quirk because Google can’t commit
Path A: Deprecated, do not dare use
Path B: Beta, do not rely
Google once again seems to have fallen into the pit of its own bureaucracy, even OpenAI looks competent by comparison.
> For Google, this is still gemini-3.1-pro-preview, right?
Flash is better than Pro for now.
Gemini has lost the race to be relevant for AI coding.
strong improvement over 3.6 flash
but luna is hard to beat @ capability / cost
For almost every section in the model card there is the message: Gemini 3.7 Flash is based on Gemini 3.6 Flash.
Same training dataset, same software, same hardware, same architecture...
I'm wondering what they changed actually for the model to be more powerful if the benchmark results are real and relevant.
Maybe just tweak settings or the reasoning prompts and called it a new version of their model?
Grok, Meta, Gemini and others all released updates to their models within around a month or two from their respective last release and made significant jumps in benchmarks all around the same time. Any guesses as to why that is? Is it just the release season and/or everyone is benchmaxxing?
its essentially the same model being trained continuously 24/7 with the company periodically publishing just a new checkpoint
each new checkpoint can benefit from better reasoning training, RL on specific tasks and more synthetic data
So why do they seem to release around the same time ? my guess is because they time major releases around quarterly earnings, investor meetings and other important business milestones. Once one company announces a major update, the others also have an incentive to ship their latest checkpoint rather than look like they r falling behind.
Sure, they are just checkpoints, that much I guess is obvious. The question is why did they not do frequent releases like this before and why are they making significant jumps in benchmarks so fast and all these companies suddenly falling into that pattern? Earning reports are not to come until end of October, that's not it.
possibly the beginning of the recursive feedback as models begin to aid in their own improvement? especially algorithmic improvements, which seems to have a lot of wide open space for gains
I'm really curious about this: the foundational paper behind today's LLMs came from Google, and some of the world's best scientists were at Google. So why are they falling so far behind in the AI race?
The "let's make money by selling/renting out TPUs" faction has won and the "let's make money by training and selling a frontier model" faction has lost.
And it's arguably not crazy, at least if SemiAnalysis's estimates are to be believed:
https://newsletter.semianalysis.com/p/gemini-is-cooked-but-g...Because they compete for the same scarce resource, the result is a resource crunch for the group that's lost: https://www.latimes.com/business/story/2026-05-18/inside-ai-...
> The "let's make money by selling/renting out TPUs" faction has won and the "let's make money by training and selling a frontier model" faction has lost.
Citation needed.
also, why can't a massive company do two things?
"why can't a massive company do two things?"
It's a variation of opportunity cost. A company that has an opportunity to take $1 and make $1.50 on it can't justify an opportunity to spend $1 and make $1.25, even though a less profitable company may make a good living on that. When considering capital allocation, Google has to consider the opportunity cost of investing more into their highly lucrative ads business. Another company that has no access to such a lucrative business uses different opportunity cost when it comes to allocating capital. It can easily be the case that Google could end up justify being in the business of renting out shovels and end up chased out of the business of using the shovels to create AIs entirely because that turns out not to be where the money is. I'm not saying that's obviously inevitable; I'm saying it's a possible and reasonable outcome.
That's why even though the industry produces giants, these giants can never just eat everything. Even though it seems like they have all the money, it isn't practical for them to try to do everything and in fact limits get hit very quickly for anything other than the primary, lucrative business.
Apparently there is no snappy term for this in the business space, according to such AI searches as I have run.
With all due respect, did you read my comment beyond the first paragraph? It addresses both points, TPU economics/pivot to sales + internal shortages making it hard to train models, to the extent they can be addressed based on public sources.
There are other factors at play, but they're more recent/second-order.
Another failed 3.5 pro run branded as 3.7 flash. It's getting sad.
Small young start-ups have to be frugal.