> Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work.
For pro mode the agents worked independently and only when they all finished did a new agent take a look at everything to merge the work into a single response. The new thing involves subagents that have been trained to cooperatively pursue a task and are allowed to communicate with each other along the way.
I tried a pro model out the other day and thought there must have been a bug in Pi’s cost calculations. But no, it’s absolutely fucking insane. Wasn’t even any better at the task.
I really suspect that the models are basically the same below, it’s all in the prompt. The way I use them, surgically, they seem to perform about the same. Fable certainly hasn’t blow my socks off.
Same. I suspect they'll get better at taking in terrible prompts over time though... Maybe that's what Fable does better, reminds me of Sora 2, it would take my crappy prompt and expound upon it. I told it once to generate a video of someone working at some company that changed its name, but the old name had historic relevance, it referred to the new company name without me telling it to, by virtue of me wanting a video of TODAY with a 90s icon.
> We generally treat GPT-5.5’s safety results as strong proxies for GPT-5.5 Pro, which is the same underlying model using a setting that makes use of parallel test time compute. As noted below, we separately evaluate GPT-5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.
I asked Claude in the browser if it could do anything like that. It wrote a little frontend app that calls the Anthropic API (with fetch()), without including a key. I expected that to fail, but it worked!
Apparently in the web chat (and also in Claude Code?[0] Though I haven't tried yet) they can call the Anthropic API and your subscription key gets auto-magicked into the requests somehow.
Those are two separate things of course (aside from the key-injection) but I guess there's no reason it couldn't run completely in the front-end... hmm...
"However, these inference optimizations, which rival Anthropic refers to as “compute multipliers,” are a big focus for all the labs. Anthropic CEO Dario Amodei has been publicly talking about the concept since at least mid-2023, when he said on a podcast that the company limits “the number of people who are aware of a given compute multiplier” because it could give other AI labs a leg up if they were to be able to replicate them. (Compute multipliers can also refer to efficiency optimizations in the model-training phase.)"
Yes, on a world with finite resources where your industry is singlehandedly siphoning ALL THE RESOURCES - hoard general efficiency optimizations and treat them as trade secrets - winning is all that matters, normal people and other species and the planet be damned.
Everything I hear about Dario these days makes me like him less and less. He sure did seem to speed run the 'tech leader with scruples' to 'tech villain' path! I guess all the cycles are compressing as we approach the singularity..
Not sure I know where I fall regarding your point: Yes to trade secrets, but also science and AI should be for the good of all.
OpenAI seems to be trading roles back with Anthropic becoming misanthropic. I hope they both start heading in the direction of how the AI field was prior to LLMs.
Collaboration and benefit for all should always be the primary motivator.
Dario tells the truth. If you look at everything through their safe AGI mission it all makes sense. They are not bs'ing about that. Also I think most people just read headlines or 10 second clips and make false extrapolations from there.
(BTW Anthropic only exists because Sam Altman is a liar, Dario admitted this.)
Like google search, this does not work because of how common long tail use is.
What you think could be a big chunk, is more likely to be a fraction of a percent of queries.
And what use is similar query caching - so you (very often! if actually cost effective, maybe half the time) get a response to a query that was different from yours. Including for when you have a lot of context input already. You’re going to get trash.
If it were constrained to only very common initial prompts, and somehow the long tail did not actually dominate as it does with Google search (can't find the reference at the moment but it was a famous article some years ago), it also wouldn't account for serious enough cost savings. Long context is what is expensive.
This might only work in constrained domains like customer service where there’s tolerance for generic answers and escalation paths. For technical work? For general purpose use, with secretly canned responses charged at full price?
Please pardon the pure speculation incoming. Yes, caching the answer doesn't seem useful. Caching the progression, the graph, may be. This is similar to making code changes with ed(1) instead of editing in vi.
The transform script(s) are cached and can be played back or adjusted. Surely for some breadth of question inputs, they map more often to similar answers--but not static answers; instead, evented edits.
It's nearly untenable for a human to keep private edit scripts to generate code changes. The extra steps for custom regex, essentially one-offs for a shared codebase, is inefficient. But maybe not to an LLM.
But there must be a ton of generic questions that people ask. Stuff like "What's the capital of country X?" - it's probably at least 10% of queries. Memories, custom instructions etc would invalidate them, but if you can return the answers basically free it's probably worth it.
Questions like that cost a tiny fraction of a cent. "What's the capital of Sri Lanka?" cost a fifth of a cent at GPT 5.5 API price, and would cost a fraction of that if the question were routed to a more suitable, cheaper model. The output was 78 tokens.
By contrast, when coding, devs typically have hundreds of thousands of tokens in the context window, and may use many millions of input tokens per day.
Caching requires the full prefix to match exactly. If a single word differs near the beginning of the prompt, nothing after that can share the cache. So this type of caching would save a few queries that cost virtually nothing, but wouldn't help with the stuff where cost matters.
Semi-related, has anyone noticed their GPT 5.5 usage in Codex being cut in half as of a couple days ago? I got a lot more mileage out of my session usage yesterday for the same workload.
I'm working in large US corporation.
And I see that I already have access to 5.6-Sol Ultra on my corporate account.
I haven't really used it yet.
2 months ago management was showing us scoreboards, praising leaders who used most tokens.
Last few weeks, we're getting weekly emails, telling us that whenever we can - we should use cheaper models, and that we should watch the page which shows our tokens usage.
Recently, I've been so eager to get new model releases in Codex. I'm hooked. I hope this accelerates development. Shows how dependant I have become to Codex.
I would assume yes - their goal is to capture consumer subscribers. Claude are going to take Fable away, and they're going to swoop in and give it to us.
Still avialable through the API. According to people that have tried both Fable nad 5.6, Fable is clearly better at coding. So i expect a lot of people to pay extra for it.
I still don't know why OpenAI doesn't put gpt-5.5-pro in Codex. It's one hell of a model and easily parallels Fable/Mythos. Sure, it'll use up your quota much faster but that's the price some users are willing to pay for absolutely high quality responses.
I think gpt-5.5-pro runs 12x parallel gpt-5.5 agents behind the scene and uses OpenAI's secret sauce to synthesize their answers into one insanely good response.
API pricing ends up being something like 20x more expensive for GPT 5.5 Pro than GPT 5.5 for actual work, even though the token cost is "only" 6x. On benchmarks where I've run both, I saw $1.12 mean per task with 5.5 and nearly $23 per task with 5.5 Pro, I guess it chews longer and harder on the problem.
If that's at all reflective of what it costs them to run it, I imagine they're in the same boat as Anthropic with Fable; they probably can't afford to offer it at subscription prices given current cost to operate it.
If 5.6 Sol Ultra has efficiency improvements (at one or more layers), and it allows OpenAI to offer a model that's competitive with Fable on the subscription plans, I'll guess a lot of folks will switch.
Fable is notably better than what came before. I watched it figure out stuff on its own over and over, on extremely hard problems, that I previously needed to guide a model to an understanding about, or work with them back and forth for several turns to figure it out together. Like, I've been reverse engineering a hardware device lately, and I've tried to tackle it a few times in the past with both some version of GPT and a couple of versions of Opus (most recently 4.7). In all cases, I barely made progress...would have gotten there eventually, probably, as I'm stubborn, but there were roadblocks constantly, with me and the model getting stumped and going around in circles in the end on every prior attempt.
Fable figured out other ways to find out what's happening, it dug into config files, found and extracted Boost-serialized data, compared that data to the observed behavior, built tools to compare the observed data with our emulated behavior, without being prompted. Would I have gotten there? Eventually, maybe. All prior models didn't; they mostly just tried the things I suggested and stopped at "well, that didn't work" or declared success after seeing results that matched their misunderstanding of the problem. I guess it possible my prior attempts with other models has "loosened the lid" on the problem; we did already have a long list of documented "this didn't work" and a pile of tools for finding out if something work. But, even so, I was impressed.
There probably will still be a "OK, let's rewrite this so it's not using lookup tables to precisely simulate the hardware behavior in software, because we don't need the noise, too" stage of the process...but, in one day with Fable, it solved a problem that I'd banged on for at least a week or too in the past with very little real progress. I don't think the models write exceedingly good code, even the best ones, but it sure does figure shit out quick.
I recently have been testing ChatGPT business at work and the quota seems to disappear almost instantly even using weaker models. Unless they dramatically increase their quotas it’ll be unusable.
I don’t know how anyone can realistically use the “business” plans - you blow through your quota so quickly. I use a consumer Pro account ($100 a month) and don’t hit the usage limits nearly as quickly. 5.5 Pro is so slow that it’s not a big deal to paste big prompts into it and come back and check on it an hour later.
Is it as good as Fable..? Fable is the first model that mostly writes without the AI slop format for me, and so I can comfortably actually copy and paste most of what it spits out.
OpenAI models have always been the worst in my experience for verbose, slop formatted responses, with each generation increasing in sloppiness.
I hate that I have had to remove it from my writing style because people assume it’s AI generated. But I think that ship has sailed. I’ll have to do without now.
For context:
> Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work.
https://openai.com/index/previewing-gpt-5-6-sol/
Can someone explain how this compares with Pro? I thought Pro was already something similar.
For pro mode the agents worked independently and only when they all finished did a new agent take a look at everything to merge the work into a single response. The new thing involves subagents that have been trained to cooperatively pursue a task and are allowed to communicate with each other along the way.
I tried a pro model out the other day and thought there must have been a bug in Pi’s cost calculations. But no, it’s absolutely fucking insane. Wasn’t even any better at the task.
I really suspect that the models are basically the same below, it’s all in the prompt. The way I use them, surgically, they seem to perform about the same. Fable certainly hasn’t blow my socks off.
> Fable certainly hasn’t blow my socks off.
Same. I suspect they'll get better at taking in terrible prompts over time though... Maybe that's what Fable does better, reminds me of Sora 2, it would take my crappy prompt and expound upon it. I told it once to generate a video of someone working at some company that changed its name, but the old name had historic relevance, it referred to the new company name without me telling it to, by virtue of me wanting a video of TODAY with a 90s icon.
Do you have a source for this, or just rumors?
The responses I get from pro don't feel like ensembles. They are often very one directional.
This can be because the summary model just picked the output from one of the sub agents.
oops
The source is the GPT 5.5 System Card:
> We generally treat GPT-5.5’s safety results as strong proxies for GPT-5.5 Pro, which is the same underlying model using a setting that makes use of parallel test time compute. As noted below, we separately evaluate GPT-5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.
https://deploymentsafety.openai.com/gpt-5-5/model-data-and-t...
There have been multiple podcasts with people from OpenAI which have confirmed this.
i would believe this will be matched with something like orchestrator-focused model: https://news.ycombinator.com/item?id=48624782
I imagine this is something like Anthropic's dynamic workflows where a JS file is created to make a little AI harness on the spot
Wow, I hadn't heard of this!
I asked Claude in the browser if it could do anything like that. It wrote a little frontend app that calls the Anthropic API (with fetch()), without including a key. I expected that to fail, but it worked!Apparently in the web chat (and also in Claude Code?[0] Though I haven't tried yet) they can call the Anthropic API and your subscription key gets auto-magicked into the requests somehow.
Those are two separate things of course (aside from the key-injection) but I guess there's no reason it couldn't run completely in the front-end... hmm...
[0] https://code.claude.com/docs/en/workflows
I wonder if it's related that that OpenAI has found a way to cut inference costs by half, according to The Information.
https://www.theinformation.com/newsletters/ai-agenda/openai-...
https://archive.ph/NEwVz
"However, these inference optimizations, which rival Anthropic refers to as “compute multipliers,” are a big focus for all the labs. Anthropic CEO Dario Amodei has been publicly talking about the concept since at least mid-2023, when he said on a podcast that the company limits “the number of people who are aware of a given compute multiplier” because it could give other AI labs a leg up if they were to be able to replicate them. (Compute multipliers can also refer to efficiency optimizations in the model-training phase.)"
Yes, on a world with finite resources where your industry is singlehandedly siphoning ALL THE RESOURCES - hoard general efficiency optimizations and treat them as trade secrets - winning is all that matters, normal people and other species and the planet be damned.
Everything I hear about Dario these days makes me like him less and less. He sure did seem to speed run the 'tech leader with scruples' to 'tech villain' path! I guess all the cycles are compressing as we approach the singularity..
Not sure I know where I fall regarding your point: Yes to trade secrets, but also science and AI should be for the good of all.
OpenAI seems to be trading roles back with Anthropic becoming misanthropic. I hope they both start heading in the direction of how the AI field was prior to LLMs.
Collaboration and benefit for all should always be the primary motivator.
He never was a leader with scruples and Anthropic has never been about anything other than grabbing power at all costs.
Dario tells the truth. If you look at everything through their safe AGI mission it all makes sense. They are not bs'ing about that. Also I think most people just read headlines or 10 second clips and make false extrapolations from there.
(BTW Anthropic only exists because Sam Altman is a liar, Dario admitted this.)
> If you look at everything through their safe AGI mission it all makes sense.
Except for, you know, all the outside investors and the forthcoming IPO.
> He sure did seem to speed run the 'tech leader with scruples' to 'tech villain' path!
What kind of rosy-eyed chump believes in the "tech leader with scruples" bullshit? It always lies.
Did some people just ignore Mark Zuckerberg and Tim Cook's sociopathy, somehow? Did anyone buy into their "privacy is a human right" nonsense?
What’s the technique? And did they buy it from thinking machines?
Maybe cache similar answers from others. Surprised if this is not already being done.
Like google search, this does not work because of how common long tail use is.
What you think could be a big chunk, is more likely to be a fraction of a percent of queries.
And what use is similar query caching - so you (very often! if actually cost effective, maybe half the time) get a response to a query that was different from yours. Including for when you have a lot of context input already. You’re going to get trash.
If it were constrained to only very common initial prompts, and somehow the long tail did not actually dominate as it does with Google search (can't find the reference at the moment but it was a famous article some years ago), it also wouldn't account for serious enough cost savings. Long context is what is expensive.
This might only work in constrained domains like customer service where there’s tolerance for generic answers and escalation paths. For technical work? For general purpose use, with secretly canned responses charged at full price?
Please pardon the pure speculation incoming. Yes, caching the answer doesn't seem useful. Caching the progression, the graph, may be. This is similar to making code changes with ed(1) instead of editing in vi.
The transform script(s) are cached and can be played back or adjusted. Surely for some breadth of question inputs, they map more often to similar answers--but not static answers; instead, evented edits.
It's nearly untenable for a human to keep private edit scripts to generate code changes. The extra steps for custom regex, essentially one-offs for a shared codebase, is inefficient. But maybe not to an LLM.
I don't understand how this fits LLM architecture at all
But there must be a ton of generic questions that people ask. Stuff like "What's the capital of country X?" - it's probably at least 10% of queries. Memories, custom instructions etc would invalidate them, but if you can return the answers basically free it's probably worth it.
Questions like that cost a tiny fraction of a cent. "What's the capital of Sri Lanka?" cost a fifth of a cent at GPT 5.5 API price, and would cost a fraction of that if the question were routed to a more suitable, cheaper model. The output was 78 tokens.
By contrast, when coding, devs typically have hundreds of thousands of tokens in the context window, and may use many millions of input tokens per day.
Caching requires the full prefix to match exactly. If a single word differs near the beginning of the prompt, nothing after that can share the cache. So this type of caching would save a few queries that cost virtually nothing, but wouldn't help with the stuff where cost matters.
I would be very surprised if they hadn’t sorted out some form of shared KV caching
I wouldn't
people really dont understand how the transformer works to think this is something trivial if possible at all
Semi-related, has anyone noticed their GPT 5.5 usage in Codex being cut in half as of a couple days ago? I got a lot more mileage out of my session usage yesterday for the same workload.
I'm working in large US corporation. And I see that I already have access to 5.6-Sol Ultra on my corporate account.
I haven't really used it yet.
2 months ago management was showing us scoreboards, praising leaders who used most tokens. Last few weeks, we're getting weekly emails, telling us that whenever we can - we should use cheaper models, and that we should watch the page which shows our tokens usage.
Recently, I've been so eager to get new model releases in Codex. I'm hooked. I hope this accelerates development. Shows how dependant I have become to Codex.
when will it be available? do we know? I don't have X, not sure if the thread mentions it.
Who cares
Related:
Previewing GPT‑5.6 Sol: a next-generation model
https://news.ycombinator.com/item?id=48689028
Will individual subscribers have access?
I would assume yes - their goal is to capture consumer subscribers. Claude are going to take Fable away, and they're going to swoop in and give it to us.
This is why I don't think Fable will be taken away. Not for long anyway.
Still avialable through the API. According to people that have tried both Fable nad 5.6, Fable is clearly better at coding. So i expect a lot of people to pay extra for it.
I love how competition is great for customers!
I still don't know why OpenAI doesn't put gpt-5.5-pro in Codex. It's one hell of a model and easily parallels Fable/Mythos. Sure, it'll use up your quota much faster but that's the price some users are willing to pay for absolutely high quality responses.
I think gpt-5.5-pro runs 12x parallel gpt-5.5 agents behind the scene and uses OpenAI's secret sauce to synthesize their answers into one insanely good response.
API pricing ends up being something like 20x more expensive for GPT 5.5 Pro than GPT 5.5 for actual work, even though the token cost is "only" 6x. On benchmarks where I've run both, I saw $1.12 mean per task with 5.5 and nearly $23 per task with 5.5 Pro, I guess it chews longer and harder on the problem.
If that's at all reflective of what it costs them to run it, I imagine they're in the same boat as Anthropic with Fable; they probably can't afford to offer it at subscription prices given current cost to operate it.
If 5.6 Sol Ultra has efficiency improvements (at one or more layers), and it allows OpenAI to offer a model that's competitive with Fable on the subscription plans, I'll guess a lot of folks will switch.
Fable is notably better than what came before. I watched it figure out stuff on its own over and over, on extremely hard problems, that I previously needed to guide a model to an understanding about, or work with them back and forth for several turns to figure it out together. Like, I've been reverse engineering a hardware device lately, and I've tried to tackle it a few times in the past with both some version of GPT and a couple of versions of Opus (most recently 4.7). In all cases, I barely made progress...would have gotten there eventually, probably, as I'm stubborn, but there were roadblocks constantly, with me and the model getting stumped and going around in circles in the end on every prior attempt.
Fable figured out other ways to find out what's happening, it dug into config files, found and extracted Boost-serialized data, compared that data to the observed behavior, built tools to compare the observed data with our emulated behavior, without being prompted. Would I have gotten there? Eventually, maybe. All prior models didn't; they mostly just tried the things I suggested and stopped at "well, that didn't work" or declared success after seeing results that matched their misunderstanding of the problem. I guess it possible my prior attempts with other models has "loosened the lid" on the problem; we did already have a long list of documented "this didn't work" and a pile of tools for finding out if something work. But, even so, I was impressed.
There probably will still be a "OK, let's rewrite this so it's not using lookup tables to precisely simulate the hardware behavior in software, because we don't need the noise, too" stage of the process...but, in one day with Fable, it solved a problem that I'd banged on for at least a week or too in the past with very little real progress. I don't think the models write exceedingly good code, even the best ones, but it sure does figure shit out quick.
I recently have been testing ChatGPT business at work and the quota seems to disappear almost instantly even using weaker models. Unless they dramatically increase their quotas it’ll be unusable.
I don’t know how anyone can realistically use the “business” plans - you blow through your quota so quickly. I use a consumer Pro account ($100 a month) and don’t hit the usage limits nearly as quickly. 5.5 Pro is so slow that it’s not a big deal to paste big prompts into it and come back and check on it an hour later.
Is it as good as Fable..? Fable is the first model that mostly writes without the AI slop format for me, and so I can comfortably actually copy and paste most of what it spits out.
OpenAI models have always been the worst in my experience for verbose, slop formatted responses, with each generation increasing in sloppiness.
> Fable is the first model that mostly writes without the AI slop format for me
I'm not that impressed by Fable's writing to be honest, still has the AI giveaways like em dash.
Humans use em dash as well.
I hate that I have had to remove it from my writing style because people assume it’s AI generated. But I think that ship has sailed. I’ll have to do without now.
Parentheses usually read better anyway.
i cant reply to hn_user2, but i have the same experience, i find myself never using emdash where i would have before
No Twitter, what’s he responding to?
https://xcancel.com/haider1/status/2073695124220006575#m
Check out LibRedirect or Predirect (MV3), it automatically redirects youtube, X, etc links to privacy-respecting frontends.
https://news.ycombinator.com/item?id=44344246
107 comments, 1 year ago.
Gamechanger..
who cares
Given that it is #1 on the front page now, people most definitely care.
Also, from the Guidelines:
> Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes.
All these names mean squat
The full conversation https://xcancel.com/haider1/status/2073695124220006575#m
The actual X link, since most of us on this platform aren't children:
https://x.com/haider1/status/2073695124220006575#m
It doesn't load. It uselessly keeps showing this in a loop:
> This process is automatic. Your browser will redirect to your requested content shortly. Please allow up to 0 second…
Remove the #m at end of the link. That's causing the redirect. Here is a cleaner link: https://xcancel.com/haider1/status/2073695124220006575
https://nitter.net/thsottiaux/status/2073933490513752151
are you, by any chance, a bot?
Not at all. I was controlling the browser myself as a human, not via any bot.