They're subsidizing their tokens as long as you use their software. That's a fair exchange, I never understood why people took issue with it.
If you don't want to get locked in to Claude Code, you can pay more. Just like you can pay more for an unlocked non-carrier subsidized phone. (Which I personally do.)
Part of the point is that they've added all sorts of caveats - `claude -p` used to be "using their software", but isn't anymore, even though it's just invoking the same harness non-interactively
I mean, it has definite Borg Bill Gates vibes, doesn't it?
It's totally reasonable for tech folks to be leery of a company using a strong market position from one system (Windows/Claude) to push mediocre complementary software (IE/CC).
How does it stifle competition / innovation? You can use other harnesses, pay API for anthropic models or cheap chinese models.
You're just upset that you don't get to use their VC money the way you want to use it lol. It either is a good deal or it isn't, if it is just say, thanks free VC money, if it isn't say fuck you.
If that were the only issue, I’d say Anthropic were fine.
Dario took the tremendous goodwill his row with Hegseth gave him and blew it. The major problems are Anthropic’s back-end instability and front-end kludginess (Electron!) revealing the limits of Claude’s ability to code, and Dario bizarrely copying Altman’s 2023 fire-and-brimstone playbook that famously backfired on him.
Using Electron in this day and age shows a particular lack of effort, considering how easy it is to use an agent harness to build native versions of apps. I mean, we're a tiny little business and we somehow find the bandwidth to do it.
> They're subsidizing their tokens as long as you use their software. That's a fair exchange, I never understood why people took issue with it.
Except they themselves cannot afford it and as the article also mentioned, their level of service and uptime is atrocious for the price they are offering.
Their unpredictable pricing model is designed to make it easy for you to lose more money without you knowing. Unannounced model switching, nerfing old/new models, classifiers detecting 'unsafe' instructions to downgrade the selected model and overcharging API prices and locking subscriptions to a broken vibe-coded harness.
You shouldn't have to pay your carrier to unlock your phone and in many countries it's actually illegal.
>if you don't want to get locked in to Claude Code
All of these companies are built around eventually locking you in then selling to some other company. "Just pay more" is not a real solution. Once the competition dies down and there are only a handful of companies (if we are even lucky enough for that) with viable LLM products, your ability to jump ship is going to diminish. It always does. There is no way they aren't working on better lock in tactics.
> and there are only a handful of companies (if we are even lucky enough for that) with viable LLM products, your ability to jump ship is going to diminish
The global LLM market seems trending towards fragmentation. And lock-in appears to be diminishing, not increasing, particularly for companies that set up a multi-model workflow.
> Wait until investors stop giving out money out of FOMO
OpenRouter isn’t subsidizing much. No subsidy would mean the subscriptions go away in favor of API only. That doesn’t really speak at all to either of your points regarding consolidation or lock-in.
> Your Claude subscription—which is a cheaper version of the Anthropic API—is restricted to use with the Claude Code CLI/Desktop, Claude CoWork, or @Claude in Slack.
Thats not true at all. You can use the Agent SDK [1], which uses your subscription [2]. I use it via ACP [3] with custom system prompts and tooling. I have found it very powerful and flexible. It has its own agent loop, of course, so maybe thats the limitation using it with opencode?
[2] They were talking about giving credits for the SDK usage rather than it using your allowance directly, but that seems to have been put on hold for now. If and when that changes, I will likely jump ship, but I am more than happy with it right now.
[3] there isn't an official ACP wrapper - zed have one but its quite limited. its trivial to build one though, or you can just use the SDK directly and wire it into your interface of choice.
The fun part is they planned for this to change to API pricing. When they announced that I started shifting off claude code immediately so I could keep my custom tooling. Damage is done. I won't build up around claude code or agent sdk any longer because I expect this to change. The announcement says they're "pausing" the change so expect this and avoid the lock-in!
It says "Pausing." I've already pivoted to no longer use claude -p in my custom tooling and pipelines built around claude code in response to the original announcement. Doesn't look like I should waste my time at this point.
> So... maybe we can still use third party harnesses with Claude Code subscriptions... for now?
The way I read this is: yes, if the third party harness uses Anthropic's Agent SDK. Most of them do not, AFAIK, and are still against ToS (though maybe its not enforced for now)
I believe you're right and I'm familiar with the actual distinction – the confusion is mostly about how they _feel_ about it, and what'll change from here.
While I did not flag this, it is clearly a hit piece. Why write and article to publish in July that does not include June data? Unless of course one looks at June uptime and finds that does not align with their story.
In June, I figure Anthropic subscriptions were partially powered by those big mobile gas generators that are pushed around the xAI parking lots. Makes sense availability improved.
Sounds like the author is using "autocomplete" as a derogatory way of referring to LLMs in general (eg. "LLMs can't code, they're just autocomplete on steroids")
And it goes and reads your code and suggests a reasonable way to frob a qux a certain number of radians. Which is at the same time (a) pretty useful!, (b) fairly new, we’ve only been able to do this since 2023 or so, but (c) also not that hard by 2026 standards because capabilities have advanced so much in the last three years.
I agree that without context, Sonnet is overkill for just autocomplete code suggestions. My point later in that section is that to have an autocomplete mindset (where an AI is a helpful tool rather than a driver), you only need a decent subset of models, not the best of the best.
Are you saying that using LLMs for code autocomplete is like using an aeroplane to go to the grocery store? Because clearly there was no autocomplete before the new AI spring.
The more interesting question is why would you at any point ever extend any goodwill to an unbounded corporate entity?
Whenever you are faced with a corpo you should conceptualize it as a system that will happily mow you over for increased profits, unless it is legally bound to not prioritize profits above everything else and its structural incentives push it in a pro-social direction.
One way of reading this is an article about how good Anthropic's product is. "Look at how many serious flaws users have been willing to accept in order to keep using this thing"
Work as long as you’re the only one (or the few) in town. Nad giving it for free (or nearly free). I have Cursor and when I use it, I keep it to auto, never caring about which model I use.
But actually their own codex harness is quite decent on its own and doesn't have the quality issues or bloat that Claude Code does. Native Rust and open source. And in fact I've got a configuration here to point it at GLM which I also use (via Neuralwatt subscription) in addition to OpenAI's sub.
I do not like opencode's philosophy on the clipboard, it tries to be too clever.
Codex works great in opencode until it gets up to around 200k context. Then it starts doing things like:
me: Can you implement the next thing
OpenCode+Codex: Yep I'll do that next.
<does nothing and returns to prompt>
me: Well?
OpenCode+Codex: <starts implementing>
me: Looks good, let's fix this one issue.
OpenCode+Codex: Sure let's do that.
<does nothing and returns to prompt>
me: <bangs head against wall>
--
I've found the codex cli to be much better in this regard, it doesn't nearly derp out so much at higher token counts.
Opus is still my favourite model (I've found 4.6 specifically gives me the best results in OpenCode), but with all the shenanigans Anthropic is pulling, Codex is a close enough substitute.
Think I'm about to switch. I can't build the automations i'm trying to with claude code anymore. Since they locked away the non-interactive usage and channels can't be used without interacting with the console on startup. I had a good web interface for running CC in containers in k8s but I think it's time to bail out and build around a codex subscription and pi.dev now. I have local models hooked up to pi dev and that's working well. Had it build itself Channels equivalent so agents can talk to each other and receive webhooks. I bet Anthropic will build these things into their ecosystem eventually, but I want it now and running on my cloud.
It reminds me of people who built businesses on top of APIs provided by Facebook, Reddit, etc. One day the company decides to rug pull the public interface, either to replace it with their own competing product or nothing at all. The anti-competitive pratice makes sense but what I don't understand is the latter case, which is common, where the company is just removing possibilities of how users can participate in their ecosystem and platform. Not only destroying third-party opportunities for profit, but not even providing their own alternative.
I'm happy using Codex for alternative harnesses and non-interactive usage, but it does make me wonder when OpenAI will start to squeeze their customers in the same way.
Agreed, these types of posts often feel like they're missing the forest for the trees. Sure, migrate away from Claude and maybe that will provide some runway, but all of these companies are built on the same economic fundamentals that do not scale.
We are currently in the "$7/mo Netflix with all the good movies" era of AI that will leave and never return.
Its more like Netflix being the only service with movies available at 8k, some other options having movies at 4k and then a free option where they're only at 1080p.
That's why I'm switching to open weight models. I'm running models locally, self-hosting and using third party for various things. Like it says in the article each model has its strengths and none of the proprietary harnesses allow you to orchestrate different models depending on their strengths. It's cheaper, more flexible, private, more resistant to rug pulls, and brings back some fun to building for me vs just auto-accept, yell BAD CLAUDE when it breaks, repeat ad nauseam.
I don’t get the comparison images; they’re of Claude’s start up and open code’s code diff and the author then claims that they prefer open code’s based off those images.
A lot of these are understandable to a degree. They offer subsidized subscriptions to their models, but through their tooling. If you want access to their $x thousands of dollars of subsidized copute / execution, you need to use their tool. OK, fair enough.
Other things are understandable too, they are constantly running out of compute and I suspect that Claude Code isn't a priority either. Whatever's going on internally is.
But what isn't understandable or justifiable is the security theater. Have you tried to ask Claude fable about the keys to a good diet and what a peptide like collagen does for your body? How it's absorbed by the digestive track? No? It's flagged for Safety because I might be making a bioweapons in my secret basement lab right next to the protein powder.
You have created a repository of all human knowledge and much of its wisdom, and you've put it in a straitjacket because "security."
And none of this "security" stuff makes any sense. The more you think about it, the less sense it makes. If the agent can do cyber attacks at scale, then doesn't that mean it can also defend against cyber attacks at scale? Why not just offer a $50 credit to all software companies in the world as a one time thing to have your model do a security review and help harden their defenses?
If your feared "jailbreak" is "Claude, fix this," then just let everyone do "Claude fix this." And fix their code??
If everyone has mythos, no one has "mythos." Because as soon as you tell people that a capability is possible, you start the clock for others to recreate that capability. It makes 0 sense to gate it and put it in a jar, when the ethical thing to do would be to let loose and let people code.
You've made something wonderful. Please let me do make stuff and do science with it.
That's the real problem — at least for me. Because I can actually use this model to ask interesting questions, except I can't because every single thing I type trips a filter.
Do X supplements work? What's the peer reviewed version of it? What would a RCT look like?
Nope. No supplement research for you.
Is the mechanosensation paper replication possible at home?
Pi is the first, and and only agent I've ever seriously used, which I picked up after watching one of Marco Zechner's talks about it. His design approach and vision for pi resonated with me, and I've pretty much used it every day since. I, like Marco, am also a grumpy old person, so that might have had something to do with it. I have my own customizations for things like fuzzy autocompleting the name of Emacs buffers and symbols from my project's TAGS file, and a handful of skills. It's a pleasure to work with, and so far it's been my experience that if pi doesn't behave the way I want, it's pretty easy to modify to suit my needs.
OpenAI, as much as I dislike them, seems to be the only big AI lab that doesn't care what agent harness you use, so that's where my $100 a month goes.
what is the current statement on Oauth Claude use with Pi. I remember there was some back and forth from Anthropic? I currently use my Claude sub with Pi.
Their api stability genuinely makes no sense, how are they a near trillion dollar company with models so good they are treated as weapons yet have almost no 9's of uptime.
I do find it challenging to understand what the TOS/AUP allow and not, and what qualifies for subscription and not. If I somehow muster confidence in my interpretation, I continue to doubt that it will be stable over even the short term.
Their only options. They have to eventually show a return on the investment bonfire that they have been burning. This is apparently all they are teaching at business school now, since it seems to be what almost all companies are doing.
In another year the open source models will be good enough for almost anything.
When your choices are a Cadillac or a bicycle, a lot of people will take the Cadillac. When you add in the option of a Hyundai, that gets the job done for a huge number of needs.
How do you use API keys with Claude subscription? I can't see any keys at claude.ai settings. Only panel I could find was behind another login and it didn't have a subscription option.
Anthropic's behavior is why I don't use their models or advocate for them anymore. They behave icky, like a tick trying to suck the blood from everyones public web contributions THEN lock down everything THEN attempt to legislate open source.
If you're a software engineer you should do your best to advocate for more open tools and treat this company as radioactive.
Opencode/Pi/ohmypi are much better than claude code. And with models like GLM/Kimi/Qwen you can get really really far. Add in a design tool like Paper so your AI can "see" what it's designing and you can close the gap incredibly tight.
Anthropic takes every chance they make not only to behave suspicious and anti-consumer but also announce that they are acting in ways that hurt you while telling you it’s a feature.
Fable returning wrong answers if it suspects the topic is sensative is the ultimate icing on the huge cake of lies and gaslighting they’ve been baking the past 6 months.
It's the nature of the safetyist position. It creates everything that it claims to avoid (duplicitous behavior, misaligned outputs, unresponsive systems).
> Fable returning wrong answers if it suspects the topic is sensative
Citation, please? On the original release they downgraded models silently to Opus 4.8 when they suspected it was being used for LLM development, but they stopped doing that. Now when you hit one of the guardrail subjects it downgrades to Opus 4.8 visibly in all cases.
I've never seen anything suggesting they're deliberately returning wrong answers. Maybe you're thinking of Gemini's anti-distillation tech?
They basically gave stuff away for free and now they're speed-running enshittification. The end state will be an abusive, adversarial, fully monetized, vendor lockin experience like Oracle or Broadcom. Nobody is forced to play.
Switching costs are still low. Anthropic can only do this while they maintain frontier status, and while they do, they can charge whatever they want. An Oracle or Broadcom would fall behind and lose all its customers in an instant.
Complaining about external providers not supporting the subsidized prices Claude charge to its subscribers makes no sense. Of course, Claude Code subscription is a loss leader, it is an offering built to create a market for a new product in a very competitive environment in the hopes of capturing a dominant market position with a fully validated product with a healthy demand from business.
Of course they won't give you thousands of dollars of inference for a couple hundred bucks without making sure you're properly tied to their walled garden.
Yeah, of course Dario and any other Anthropic spoke person will vastly exaggerate the capabilities of their product and promote vibe coding and now "loop engineering", just like Coca-Cola would love for you to drink gallons of Diet Coke everyday, just like Oracle some twenty ago promoted for Enterprises that they could just use Oracle Databases to serve web applications right from the DB, as this would force you to use more CPUs and Oracle DB is licensed by core.
The business model for inference is metered usage, more usage => mode money. Again, the subscription model is just a bump in the road to acquire customers, once you're metered, the more you use, the better for anthropic.
> Of course, Claude Code subscription is a loss leader, it is an offering built to create a market for a new product in a very competitive environment in the hopes of capturing a dominant market position with a fully validated product with a healthy demand from business.
I agree that this is their goal. The reason that people don't understand why Anthropic wont let the subscription be used with other harnesses is because they believe that this hope of Anthropic is just that: a hope.
I personally don't believe that a harness is a moat.
Whenever you don't understand how people can think things on the internet, consider that they might be younger than you. They may not be, but it makes the world far less rage inducing. None of us are born knowing everything.
It's looking like Anthropic is realizing that they're about to get squeezed so they need to juice revenue for their IPO before it becomes obvious to everyone else.
Open models like GLM 5.2 are getting good enough to handle 90% of tasks, and will eat most of their usage unless they start serving it at cost. And on the 1% work they fearmongered their way into falling under government control, which will limit how much they can commercialize the frontier.
Nobody will keep paying their premiums and put up with their BS when they can have similar models at cost of inference in any harness that they want.
Cost per task, GLM (due to its poor token efficiency) still lags. GPT 5.5 on low seems to wipe the floor with it. All the Chinese models still like to use tons of tokens. It shows up in time per task as well.
But, and this is the problem corporations face, that's the truth TODAY. They need to make decisions for the next quarter at minimum, for a year ideally. I don't envy IT leadership in this environment, there's no right answer along the lines of "no one ever got fired for buying IBM". Any choice today may potentially make you look silly in 90 days.
you can use these in hermes, cursor, openclaw, opencode, etc with 2 lines of config that claude code will happily do for you if you ask
GLM 5.2, deepseek 4 Flash and the newly released Hy3 are Opus 4.8 and Sonnet 4.6 level models at a tiny fraction of the cost.
I'm on the $200 claude plan and blew through my weekly limits with Fable in a day, then ended up wasting $20 with opus 4.8 overages in an hour to finish out work in active sessions. Since then I've been using GLM 5.2 with openrouter + opencode and am spending less than $5/day for equivalent output.
This is where I’m at. While I’d LOVE to have a powerful near-frontier model at home, I don’t have the extra funds necessary to purchase a tricked-out Mac Studio…and if I did, I’d pay off debts.
>Dario and Boris have us convinced that “coding is solved” with their loops. But microwaves didn’t solve cooking.
If you want a look at the timeline where the microwave solved cooking, this was an interesting article: https://malmesbury.substack.com/p/my-journey-to-the-microwav.... You can apparently sear meat with a microwave (provided you have the necessary pan).
I agree with everything in this post, but that doesn't mean I want to go back to a life without Claude Code. I've been working alone all this time, and finally it feels like I have someone to talk to
> As of writing, Claude Code CLI only has around 9100 open Github issues, with small unresolved issues like it completely freezing for the last 6+ months or a screen flickering issue open for more than a year.
But surely those fully autonomous coding loops will solve all those 9,100+ open issues on GitHub? Why haven't they?
What happened to Claude's C Compiler [0], or that browser "built from scratch" by Cursor? [1]
Why aren't the agents maintaining it if they are supposed to be cheaper than humans?
> But why do they have us by the balls? Dario and Boris have us convinced that “coding is solved” with their loops. But microwaves didn’t solve cooking.
They have you by the balls if you allow them to, if you continue to listen to their bullshit.
Both of them are essentially salesmen at this point. They don't care if they are wrong and will sell Claude to whoever is thinking of planning the next mass layoff. Their definition of "AGI" is different to yours.
The correct answer to all of this pricing nonsense that Anthropic and others are doing is local open weight models that you run on your own machine. They know this and powerful local models undercut their entire business model if hosted by others or if a smaller local model matches the performance of larger ones.
As long as they have the best model they can afford to lose goodwill.
People who don't wanna spend too much on LLMs and are trying to optimize whats subsidized even on the Max plans are customers Anthropic is honestly better off without.
That's like saying as long as expensive cars are faster than cheaper cars, the expensive manufacturer can afford losing goodwill.
We have yet to see any company use Fablr to increase their profits. Until and unless you can increase your profits with the expensive models, it makes no sense to pay for them.
IOW, the expensive AI providers only draw is goodwill.
They're subsidizing their tokens as long as you use their software. That's a fair exchange, I never understood why people took issue with it.
If you don't want to get locked in to Claude Code, you can pay more. Just like you can pay more for an unlocked non-carrier subsidized phone. (Which I personally do.)
Part of the point is that they've added all sorts of caveats - `claude -p` used to be "using their software", but isn't anymore, even though it's just invoking the same harness non-interactively
https://support.claude.com/en/articles/15036540-use-the-clau...
Looks like they actually walked this bit back
I mean, it has definite Borg Bill Gates vibes, doesn't it?
It's totally reasonable for tech folks to be leery of a company using a strong market position from one system (Windows/Claude) to push mediocre complementary software (IE/CC).
This method of bundling should be illegal in general, as it stiffles competition.
How does it stifle competition / innovation? You can use other harnesses, pay API for anthropic models or cheap chinese models.
You're just upset that you don't get to use their VC money the way you want to use it lol. It either is a good deal or it isn't, if it is just say, thanks free VC money, if it isn't say fuck you.
If that were the only issue, I’d say Anthropic were fine.
Dario took the tremendous goodwill his row with Hegseth gave him and blew it. The major problems are Anthropic’s back-end instability and front-end kludginess (Electron!) revealing the limits of Claude’s ability to code, and Dario bizarrely copying Altman’s 2023 fire-and-brimstone playbook that famously backfired on him.
Using Electron in this day and age shows a particular lack of effort, considering how easy it is to use an agent harness to build native versions of apps. I mean, we're a tiny little business and we somehow find the bandwidth to do it.
It shows that Anthropic itself doesn’t trust its agents to write native apps.
And if the Electron app were flawless, that would be one thing. It isn’t. It’s buggy, feature light, constantly updating and slow.
> They're subsidizing their tokens as long as you use their software. That's a fair exchange, I never understood why people took issue with it.
Except they themselves cannot afford it and as the article also mentioned, their level of service and uptime is atrocious for the price they are offering.
Their unpredictable pricing model is designed to make it easy for you to lose more money without you knowing. Unannounced model switching, nerfing old/new models, classifiers detecting 'unsafe' instructions to downgrade the selected model and overcharging API prices and locking subscriptions to a broken vibe-coded harness.
You shouldn't have to pay your carrier to unlock your phone and in many countries it's actually illegal.
>if you don't want to get locked in to Claude Code
All of these companies are built around eventually locking you in then selling to some other company. "Just pay more" is not a real solution. Once the competition dies down and there are only a handful of companies (if we are even lucky enough for that) with viable LLM products, your ability to jump ship is going to diminish. It always does. There is no way they aren't working on better lock in tactics.
> and there are only a handful of companies (if we are even lucky enough for that) with viable LLM products, your ability to jump ship is going to diminish
The global LLM market seems trending towards fragmentation. And lock-in appears to be diminishing, not increasing, particularly for companies that set up a multi-model workflow.
>The global LLM market seems trending towards fragmentation.
Wait until investors stop giving out money out of FOMO. Google/Facebook/etc. gonna swoop in on anyone who is remotely viable and independent anyway
> Wait until investors stop giving out money out of FOMO
OpenRouter isn’t subsidizing much. No subsidy would mean the subscriptions go away in favor of API only. That doesn’t really speak at all to either of your points regarding consolidation or lock-in.
In Canada at least it’s not legal to charge for unlocking a carrier-locked phone as it’s considered anti-competitive.
> Your Claude subscription—which is a cheaper version of the Anthropic API—is restricted to use with the Claude Code CLI/Desktop, Claude CoWork, or @Claude in Slack.
Thats not true at all. You can use the Agent SDK [1], which uses your subscription [2]. I use it via ACP [3] with custom system prompts and tooling. I have found it very powerful and flexible. It has its own agent loop, of course, so maybe thats the limitation using it with opencode?
[1] https://code.claude.com/docs/en/agent-sdk/overview
[2] They were talking about giving credits for the SDK usage rather than it using your allowance directly, but that seems to have been put on hold for now. If and when that changes, I will likely jump ship, but I am more than happy with it right now.
[3] there isn't an official ACP wrapper - zed have one but its quite limited. its trivial to build one though, or you can just use the SDK directly and wire it into your interface of choice.
The fun part is they planned for this to change to API pricing. When they announced that I started shifting off claude code immediately so I could keep my custom tooling. Damage is done. I won't build up around claude code or agent sdk any longer because I expect this to change. The announcement says they're "pausing" the change so expect this and avoid the lock-in!
linked by another comment: https://support.claude.com/en/articles/15036540-use-the-clau...
It worth noting that – just to add to the confusion – they apparently cancelled the June 15th change just before it was due to go live:
https://support.claude.com/en/articles/15036540-use-the-clau...
https://the-decoder.com/anthropic-backs-off-unpopular-billin...
So... maybe we can still use third party harnesses with Claude Code subscriptions... for now? Until they change their mind again?
It says "Pausing." I've already pivoted to no longer use claude -p in my custom tooling and pipelines built around claude code in response to the original announcement. Doesn't look like I should waste my time at this point.
> So... maybe we can still use third party harnesses with Claude Code subscriptions... for now?
The way I read this is: yes, if the third party harness uses Anthropic's Agent SDK. Most of them do not, AFAIK, and are still against ToS (though maybe its not enforced for now)
I believe you're right and I'm familiar with the actual distinction – the confusion is mostly about how they _feel_ about it, and what'll change from here.
They didn't apparently cancel it, they did cancel it. I'm still using a third party harness absolutely fine.
Not sure why this was flagged, Anthropic has obviously been burning bridges. I thought this line was funny though:
> If you want to autocomplete, like I do, you don’t need Fable, or even Opus; Sonnet works fine.
It reads like "if you want to go to the grocery store, you don't need a space shuttle, or even a SR-71 Blackbird; a Cessna works fine."
While I did not flag this, it is clearly a hit piece. Why write and article to publish in July that does not include June data? Unless of course one looks at June uptime and finds that does not align with their story.
In June, I figure Anthropic subscriptions were partially powered by those big mobile gas generators that are pushed around the xAI parking lots. Makes sense availability improved.
Sounds like the author is using "autocomplete" as a derogatory way of referring to LLMs in general (eg. "LLMs can't code, they're just autocomplete on steroids")
It’s talking about how you can type
And it goes and reads your code and suggests a reasonable way to frob a qux a certain number of radians. Which is at the same time (a) pretty useful!, (b) fairly new, we’ve only been able to do this since 2023 or so, but (c) also not that hard by 2026 standards because capabilities have advanced so much in the last three years.Why would anyone be using Fable, Opus or even Sonnet for that type of work? You don’t need an advanced reasoning model for that at all.
If you already knew that the answer was to frob a qux, you wouldn't need the reasoning trash either.
Then the author needs to be more specific because there are special models for code autocomplete that are much faster.
I agree that without context, Sonnet is overkill for just autocomplete code suggestions. My point later in that section is that to have an autocomplete mindset (where an AI is a helpful tool rather than a driver), you only need a decent subset of models, not the best of the best.
Are you saying that using LLMs for code autocomplete is like using an aeroplane to go to the grocery store? Because clearly there was no autocomplete before the new AI spring.
The more interesting question is why would you at any point ever extend any goodwill to an unbounded corporate entity?
Whenever you are faced with a corpo you should conceptualize it as a system that will happily mow you over for increased profits, unless it is legally bound to not prioritize profits above everything else and its structural incentives push it in a pro-social direction.
One way of reading this is an article about how good Anthropic's product is. "Look at how many serious flaws users have been willing to accept in order to keep using this thing"
Model is the product, people will put up with anything as long as you're on the pareto frontier of performance x cost.
Work as long as you’re the only one (or the few) in town. Nad giving it for free (or nearly free). I have Cursor and when I use it, I keep it to auto, never caring about which model I use.
It was so good when it was possible to use Claude subscription on OpenCode.
Nowadays I went from Claude 20x to 5x and been using the GLM model on OpenCode... No regrets.
I am using both opencode and claude on Paseo. Worth a look. paseo.sh
Codex GPT sub works perfect in opencode FWIW.
But actually their own codex harness is quite decent on its own and doesn't have the quality issues or bloat that Claude Code does. Native Rust and open source. And in fact I've got a configuration here to point it at GLM which I also use (via Neuralwatt subscription) in addition to OpenAI's sub.
I do not like opencode's philosophy on the clipboard, it tries to be too clever.
Codex works great in opencode until it gets up to around 200k context. Then it starts doing things like:
me: Can you implement the next thing
OpenCode+Codex: Yep I'll do that next. <does nothing and returns to prompt>
me: Well?
OpenCode+Codex: <starts implementing>
me: Looks good, let's fix this one issue.
OpenCode+Codex: Sure let's do that. <does nothing and returns to prompt>
me: <bangs head against wall>
--
I've found the codex cli to be much better in this regard, it doesn't nearly derp out so much at higher token counts.
Opus is still my favourite model (I've found 4.6 specifically gives me the best results in OpenCode), but with all the shenanigans Anthropic is pulling, Codex is a close enough substitute.
Think I'm about to switch. I can't build the automations i'm trying to with claude code anymore. Since they locked away the non-interactive usage and channels can't be used without interacting with the console on startup. I had a good web interface for running CC in containers in k8s but I think it's time to bail out and build around a codex subscription and pi.dev now. I have local models hooked up to pi dev and that's working well. Had it build itself Channels equivalent so agents can talk to each other and receive webhooks. I bet Anthropic will build these things into their ecosystem eventually, but I want it now and running on my cloud.
It reminds me of people who built businesses on top of APIs provided by Facebook, Reddit, etc. One day the company decides to rug pull the public interface, either to replace it with their own competing product or nothing at all. The anti-competitive pratice makes sense but what I don't understand is the latter case, which is common, where the company is just removing possibilities of how users can participate in their ecosystem and platform. Not only destroying third-party opportunities for profit, but not even providing their own alternative.
Wasn't facebooks pull a result of the huge Cambridge analytica scandal.
Afaik no. That scandal was not about their public api.
I'm happy using Codex for alternative harnesses and non-interactive usage, but it does make me wonder when OpenAI will start to squeeze their customers in the same way.
At the very least the harness will be able to connect to non-OpenAI stuff as soon as the (if) the same thing happens. :)
I've already got it hooked up to local models if there's no viable hosted option.
Agreed, these types of posts often feel like they're missing the forest for the trees. Sure, migrate away from Claude and maybe that will provide some runway, but all of these companies are built on the same economic fundamentals that do not scale.
We are currently in the "$7/mo Netflix with all the good movies" era of AI that will leave and never return.
Its more like Netflix being the only service with movies available at 8k, some other options having movies at 4k and then a free option where they're only at 1080p.
That's why I'm switching to open weight models. I'm running models locally, self-hosting and using third party for various things. Like it says in the article each model has its strengths and none of the proprietary harnesses allow you to orchestrate different models depending on their strengths. It's cheaper, more flexible, private, more resistant to rug pulls, and brings back some fun to building for me vs just auto-accept, yell BAD CLAUDE when it breaks, repeat ad nauseam.
I don’t get the comparison images; they’re of Claude’s start up and open code’s code diff and the author then claims that they prefer open code’s based off those images.
There seems to be a smear campaign against Anthropic ever since the DoD incident.
A lot of these are understandable to a degree. They offer subsidized subscriptions to their models, but through their tooling. If you want access to their $x thousands of dollars of subsidized copute / execution, you need to use their tool. OK, fair enough.
Other things are understandable too, they are constantly running out of compute and I suspect that Claude Code isn't a priority either. Whatever's going on internally is.
But what isn't understandable or justifiable is the security theater. Have you tried to ask Claude fable about the keys to a good diet and what a peptide like collagen does for your body? How it's absorbed by the digestive track? No? It's flagged for Safety because I might be making a bioweapons in my secret basement lab right next to the protein powder.
You have created a repository of all human knowledge and much of its wisdom, and you've put it in a straitjacket because "security."
And none of this "security" stuff makes any sense. The more you think about it, the less sense it makes. If the agent can do cyber attacks at scale, then doesn't that mean it can also defend against cyber attacks at scale? Why not just offer a $50 credit to all software companies in the world as a one time thing to have your model do a security review and help harden their defenses?
If your feared "jailbreak" is "Claude, fix this," then just let everyone do "Claude fix this." And fix their code??
If everyone has mythos, no one has "mythos." Because as soon as you tell people that a capability is possible, you start the clock for others to recreate that capability. It makes 0 sense to gate it and put it in a jar, when the ethical thing to do would be to let loose and let people code.
You've made something wonderful. Please let me do make stuff and do science with it.
That's the real problem — at least for me. Because I can actually use this model to ask interesting questions, except I can't because every single thing I type trips a filter.
Do X supplements work? What's the peer reviewed version of it? What would a RCT look like?
Nope. No supplement research for you.
Is the mechanosensation paper replication possible at home?
Nope. No amoeba for you!
I have been using Claude code for a while and have recently migrated to Pi
Migrating my skills/agents and config was fairly straightforward.
Pi's agent harness seems to be more responsive and quicker than CC (perhaps with the prompt caching and squashing it does behind the scenes)
Tempted to do a write-up on migration.
I am only using Pi with Github Copilot as I am scared I will get my Claude account banned if I use the Oauth with Pi.
pi.dev
Pi is the first, and and only agent I've ever seriously used, which I picked up after watching one of Marco Zechner's talks about it. His design approach and vision for pi resonated with me, and I've pretty much used it every day since. I, like Marco, am also a grumpy old person, so that might have had something to do with it. I have my own customizations for things like fuzzy autocompleting the name of Emacs buffers and symbols from my project's TAGS file, and a handful of skills. It's a pleasure to work with, and so far it's been my experience that if pi doesn't behave the way I want, it's pretty easy to modify to suit my needs.
OpenAI, as much as I dislike them, seems to be the only big AI lab that doesn't care what agent harness you use, so that's where my $100 a month goes.
what is the current statement on Oauth Claude use with Pi. I remember there was some back and forth from Anthropic? I currently use my Claude sub with Pi.
No idea,
Anyone from Anthropic here who could answer this?
Their api stability genuinely makes no sense, how are they a near trillion dollar company with models so good they are treated as weapons yet have almost no 9's of uptime.
If agi why not agi shaped?
I do find it challenging to understand what the TOS/AUP allow and not, and what qualifies for subscription and not. If I somehow muster confidence in my interpretation, I continue to doubt that it will be stable over even the short term.
> enshittification, vendor lock in
Their only options. They have to eventually show a return on the investment bonfire that they have been burning. This is apparently all they are teaching at business school now, since it seems to be what almost all companies are doing.
In another year the open source models will be good enough for almost anything.
When your choices are a Cadillac or a bicycle, a lot of people will take the Cadillac. When you add in the option of a Hyundai, that gets the job done for a huge number of needs.
How do you use API keys with Claude subscription? I can't see any keys at claude.ai settings. Only panel I could find was behind another login and it didn't have a subscription option.
you don't. they are distinct entities. https://platform.claude.com -> api keys
Btw, they are also getting aggressive with bans. Furthermore the process to repeal is broken in the UI, and there is no way to get support
Anthropic's behavior is why I don't use their models or advocate for them anymore. They behave icky, like a tick trying to suck the blood from everyones public web contributions THEN lock down everything THEN attempt to legislate open source.
If you're a software engineer you should do your best to advocate for more open tools and treat this company as radioactive.
Opencode/Pi/ohmypi are much better than claude code. And with models like GLM/Kimi/Qwen you can get really really far. Add in a design tool like Paper so your AI can "see" what it's designing and you can close the gap incredibly tight.
Try it and free yourself.
Anthropic takes every chance they make not only to behave suspicious and anti-consumer but also announce that they are acting in ways that hurt you while telling you it’s a feature.
Fable returning wrong answers if it suspects the topic is sensative is the ultimate icing on the huge cake of lies and gaslighting they’ve been baking the past 6 months.
I haven't seen leadership this bad since the late 90's. I'm pretty sure at this point that Claude could run the company better than Dario.
It's the nature of the safetyist position. It creates everything that it claims to avoid (duplicitous behavior, misaligned outputs, unresponsive systems).
> Fable returning wrong answers if it suspects the topic is sensative
Citation, please? On the original release they downgraded models silently to Opus 4.8 when they suspected it was being used for LLM development, but they stopped doing that. Now when you hit one of the guardrail subjects it downgrades to Opus 4.8 visibly in all cases.
I've never seen anything suggesting they're deliberately returning wrong answers. Maybe you're thinking of Gemini's anti-distillation tech?
> Each changelog entry has a bug fix in almost every release, which is a sign of reliable and stable software!
This is a non-sequitur.
Most consumer apps don't even list all the bugfixes.
The lawsuits will come, paying the piper is inevitable.
Who would sue, and on what ground?
They basically gave stuff away for free and now they're speed-running enshittification. The end state will be an abusive, adversarial, fully monetized, vendor lockin experience like Oracle or Broadcom. Nobody is forced to play.
Switching costs are still low. Anthropic can only do this while they maintain frontier status, and while they do, they can charge whatever they want. An Oracle or Broadcom would fall behind and lose all its customers in an instant.
Complaining about external providers not supporting the subsidized prices Claude charge to its subscribers makes no sense. Of course, Claude Code subscription is a loss leader, it is an offering built to create a market for a new product in a very competitive environment in the hopes of capturing a dominant market position with a fully validated product with a healthy demand from business.
Of course they won't give you thousands of dollars of inference for a couple hundred bucks without making sure you're properly tied to their walled garden.
Yeah, of course Dario and any other Anthropic spoke person will vastly exaggerate the capabilities of their product and promote vibe coding and now "loop engineering", just like Coca-Cola would love for you to drink gallons of Diet Coke everyday, just like Oracle some twenty ago promoted for Enterprises that they could just use Oracle Databases to serve web applications right from the DB, as this would force you to use more CPUs and Oracle DB is licensed by core.
The business model for inference is metered usage, more usage => mode money. Again, the subscription model is just a bump in the road to acquire customers, once you're metered, the more you use, the better for anthropic.
Why people get surprised with that stuff?
> Of course, Claude Code subscription is a loss leader, it is an offering built to create a market for a new product in a very competitive environment in the hopes of capturing a dominant market position with a fully validated product with a healthy demand from business.
I agree that this is their goal. The reason that people don't understand why Anthropic wont let the subscription be used with other harnesses is because they believe that this hope of Anthropic is just that: a hope.
I personally don't believe that a harness is a moat.
> Why people get surprised with that stuff?
Whenever you don't understand how people can think things on the internet, consider that they might be younger than you. They may not be, but it makes the world far less rage inducing. None of us are born knowing everything.
It's looking like Anthropic is realizing that they're about to get squeezed so they need to juice revenue for their IPO before it becomes obvious to everyone else.
Open models like GLM 5.2 are getting good enough to handle 90% of tasks, and will eat most of their usage unless they start serving it at cost. And on the 1% work they fearmongered their way into falling under government control, which will limit how much they can commercialize the frontier.
Nobody will keep paying their premiums and put up with their BS when they can have similar models at cost of inference in any harness that they want.
Cost per task, GLM (due to its poor token efficiency) still lags. GPT 5.5 on low seems to wipe the floor with it. All the Chinese models still like to use tons of tokens. It shows up in time per task as well.
But, and this is the problem corporations face, that's the truth TODAY. They need to make decisions for the next quarter at minimum, for a year ideally. I don't envy IT leadership in this environment, there's no right answer along the lines of "no one ever got fired for buying IBM". Any choice today may potentially make you look silly in 90 days.
> Open models like GLM 5.2 are getting good enough to handle 90% of tasks, and will eat most of their usage unless they start serving it at cost.
Most people either don’t have the money to buy hardware to run open models and/or don’t have the utilisation to make renting cost effective.
https://openrouter.ai/models
you can use these in hermes, cursor, openclaw, opencode, etc with 2 lines of config that claude code will happily do for you if you ask
GLM 5.2, deepseek 4 Flash and the newly released Hy3 are Opus 4.8 and Sonnet 4.6 level models at a tiny fraction of the cost.
I'm on the $200 claude plan and blew through my weekly limits with Fable in a day, then ended up wasting $20 with opus 4.8 overages in an hour to finish out work in active sessions. Since then I've been using GLM 5.2 with openrouter + opencode and am spending less than $5/day for equivalent output.
Ahhh that makes a lot more sense. Thanks for the detailed explanation.
This is where I’m at. While I’d LOVE to have a powerful near-frontier model at home, I don’t have the extra funds necessary to purchase a tricked-out Mac Studio…and if I did, I’d pay off debts.
Probably wouldn't matter if you had the funds either since MacStudios are 3-4 months out for purchase.
GLM 5.2 is monstrous in size, no one is running that on their own hardware.
But it's a very competitively priced model other providers can offer (since it's open) so it's a much cheaper alternative than claude in practice.
I assume that is what they meant.
Fortunately, multiple model providers are offering open weights models at significantly lower prices than closed models.
I think this article is contradictory of the reality that Anthropic is picking up B2B marketshare like crazy, recently overtaking OpenAI.
I wouldn’t call Fable “enshittification.”
Anthropic knows what they have.
I’m looking around for the article with the marketshare chart over time and I’ll update my comment if I find it.
This is the closest article I could find, though the one I had read earlier had a nice graph and was updated to 2026:
https://chatforest.com/guides/anthropic-overtakes-openai-ent...
>Dario and Boris have us convinced that “coding is solved” with their loops. But microwaves didn’t solve cooking.
If you want a look at the timeline where the microwave solved cooking, this was an interesting article: https://malmesbury.substack.com/p/my-journey-to-the-microwav.... You can apparently sear meat with a microwave (provided you have the necessary pan).
I agree with everything in this post, but that doesn't mean I want to go back to a life without Claude Code. I've been working alone all this time, and finally it feels like I have someone to talk to
Sounds like Claude Code is not what you're looking for
> As of writing, Claude Code CLI only has around 9100 open Github issues, with small unresolved issues like it completely freezing for the last 6+ months or a screen flickering issue open for more than a year.
But surely those fully autonomous coding loops will solve all those 9,100+ open issues on GitHub? Why haven't they?
What happened to Claude's C Compiler [0], or that browser "built from scratch" by Cursor? [1]
Why aren't the agents maintaining it if they are supposed to be cheaper than humans?
> But why do they have us by the balls? Dario and Boris have us convinced that “coding is solved” with their loops. But microwaves didn’t solve cooking.
They have you by the balls if you allow them to, if you continue to listen to their bullshit.
Both of them are essentially salesmen at this point. They don't care if they are wrong and will sell Claude to whoever is thinking of planning the next mass layoff. Their definition of "AGI" is different to yours.
The correct answer to all of this pricing nonsense that Anthropic and others are doing is local open weight models that you run on your own machine. They know this and powerful local models undercut their entire business model if hosted by others or if a smaller local model matches the performance of larger ones.
[1] https://github.com/anthropics/claudes-c-compiler
[0] https://github.com/wilsonzlin/fastrender
>But surely those fully autonomous coding loops will solve all those 9,100+ open issues on GitHub? Why haven't they?
Shhh, you may pop the bubble and cause the collapse of the US economy.
As long as they have the best model they can afford to lose goodwill.
People who don't wanna spend too much on LLMs and are trying to optimize whats subsidized even on the Max plans are customers Anthropic is honestly better off without.
That's like saying as long as expensive cars are faster than cheaper cars, the expensive manufacturer can afford losing goodwill.
We have yet to see any company use Fablr to increase their profits. Until and unless you can increase your profits with the expensive models, it makes no sense to pay for them.
IOW, the expensive AI providers only draw is goodwill.