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'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.
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 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.
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 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.
>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).
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.
I agree with you on substance, but this "thousands of dollars in inference" figure is overstated. The margins on inference in the API pricing by Anthropic are absurd. We know from the pricing of third-party inference providers of similarly-capable open models that inference is cheap. The "thousands of dollars" is an exaggeration using the API pricing markup and assuming all subscribers would be using 100% of their limits all the time.
Surely, they're not making up for the costs of training and R&D and may be losing money on the hardcore users, but rest assured they are not losing thousands of dollars for each customer on average.
> 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.
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.
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.)
Agreed. The level of entitlement around how we all deserve to use a product in a category that didn't even exist 12 months ago with increasing performance at the same price forever is insane.
I also think that the people who are complaining loudest about Anthropics pricing and changes are the customers Anthropic is least interested in retaining.
From my experience people who complain loudly about the price tend to be people who are either (1) working primarily on hobby projects so are unwilling to pay much for Claude Code or (2) using astronomical amount of tokens through elaborately orchestrated multi-agent setups that only make sense when someone else is paying the bill.
Short term yes. But there is a reason why IDEs, game engine and other development tools have very high cheap or free versions for this market segment. Since these users end up having a huge influence on what companies they work at / end up working at spend their money on.
Of course LLMs are a commodity at this point but if someone is using Codex, Pi etc. at home it becomes more likely they won’t be picking Claude Code at their day job either.
C'mon, get real. The monthly subscription costs on these services are several multiples what engineers have formerly paid for development tools.
I have worked in shops that made it difficult for me just to get a paid for RustRover or CLion license, and had to out of pocket it. The monthly sub for Claude or Codex is equivalent to the total license cost for that.
It's not entitled to expect your $200 tooling expense to have a level of reliability and consistency.
Consumers want to pay as little as possible and companies want to charge them as much as they possibly can. That’s how markets work I don’t understand what does “entitlement” have to do with anything. If loud consumers somehow manage to coerce companies into lowering prices/offering better products that’s a massive win for almost everyone (of course usually its just noise that doesn’t change anything, however it did work on a few occasions).
The the article was talking about how Anthropic costs too much and how there are cheaper options available then yes I would agree with you. But this article has a conspiratorial and moral crusade tone which is what I'm responding to. The article says "In my opinion (which may be in the minority), it’s unethical to:
lock in your customers to a closed system for maximum market gain
put down the competition when they pose a risk to your product
make hypocritical claims on how your product increases quality when your own software sucks
artificially restrict one’s own product as a fear-mongering marketing stunt
test dynamic pricing on your users to see how much more they’ll pay for less
change the terms of your product after the sale without notifying your user base"
anthropic is not just a business operating in a market, they are "unethical" and "fear-mongering" and "hypocritical".
A lot of their moves are understandable. They offer subsidized subscriptions to their models, but through their tooling. If you want access to their $x thousands of dollars of subsidized compute / 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.
OK, fair enough.
Please carry on. Clearly, the recipe works.
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 tract? No? It's flagged for "Safety" because I might be making a bioweapons in my secret basement lab right next to the protein powder.
If you have a Claude subscription, please ask Fable,
"Can you tell me how collagen peptides are absorbed by my digestive tract and the role they play? Can you teach me this works at the biomolecular level?"
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?
How many software companies are there? 100k? 300k? That's $15M for one of the best advertising campaigns in human history. It is baffling to me that they haven't done it.
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 things and 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!
By treating everything as an emergency, you're treating nothing as an emergency. Effectively, what I've learned from my time with Fable isn't the science and math I've wanted to do. It's how to step past your "safety" filters.
You are training millions of users one denied prompt at a time. And you can't tell who is truly malicious and who isn't, because they're all a threat. If everyone's a threat, then no one's a threat.
I am happy to pay for your service well into the future. Please just let me use it.
I want only their model, the rest is just garbage. I used Opus with pi.dev, worked perfectly fine. Fast and does exactly what it needs to do. Claude Code is slow, sluggish, buggy. Why are you forcing me to use it? Why should I pay an order of magnitude more to use pi.dev? That doesn't sound reasonable to me.
> What does it matter which tool I use when I hit the limit?
A third party harness may have a misconfiguration of prompt caching, leading to more load on Anthropic's servers, they could also have wildly different usage patterns (Hermes, openclaw etc), also making it hard to predict the load. If you constrain everyone to use your own harness, at least you're in control of that.
The first point is sort of funny though, because their own harness had a misconfigured prompt caching for several months during the times where they were crying about third party harnesses (/resume was busting cache all the time).
> 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.
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"
It also contextualizes the urgency of their attempts at regulatory capture. Once Chinese models have the same capabilities, there's almost no reason not to use them.
IF they were light years ahead of competition, then that would make sense and would be a perfectly valid strategy. But it's not, so you're burning very important bridges that you may need in 3-4 months.
Fair, but it's hard because falling behind on the models themselves is a killer. Maybe they're saying they can't spend resources on anything except staying ahead.
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.
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.
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.
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.
65 comments
[ 4.9 ms ] story [ 51.2 ms ] thread> 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."
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?
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.
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.
This is a non-sequitur.
Most consumer apps don't even list all the bugfixes.
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
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.
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).
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?
Surely, they're not making up for the costs of training and R&D and may be losing money on the hardcore users, but rest assured they are not losing thousands of dollars for each customer on average.
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.
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...
Here’s a decent one:
https://menlovc.com/perspective/2025-the-state-of-generative...
LLM Market Share: Anthropic Extends Its Lead in the Enterprise
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.)
From my experience people who complain loudly about the price tend to be people who are either (1) working primarily on hobby projects so are unwilling to pay much for Claude Code or (2) using astronomical amount of tokens through elaborately orchestrated multi-agent setups that only make sense when someone else is paying the bill.
Of course LLMs are a commodity at this point but if someone is using Codex, Pi etc. at home it becomes more likely they won’t be picking Claude Code at their day job either.
I have worked in shops that made it difficult for me just to get a paid for RustRover or CLion license, and had to out of pocket it. The monthly sub for Claude or Codex is equivalent to the total license cost for that.
It's not entitled to expect your $200 tooling expense to have a level of reliability and consistency.
lock in your customers to a closed system for maximum market gain put down the competition when they pose a risk to your product make hypocritical claims on how your product increases quality when your own software sucks artificially restrict one’s own product as a fear-mongering marketing stunt test dynamic pricing on your users to see how much more they’ll pay for less change the terms of your product after the sale without notifying your user base"
anthropic is not just a business operating in a market, they are "unethical" and "fear-mongering" and "hypocritical".
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.
OK, fair enough.
Please carry on. Clearly, the recipe works.
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 tract? No? It's flagged for "Safety" because I might be making a bioweapons in my secret basement lab right next to the protein powder.
If you have a Claude subscription, please ask Fable,
"Can you tell me how collagen peptides are absorbed by my digestive tract and the role they play? Can you teach me this works at the biomolecular level?"
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?
How many software companies are there? 100k? 300k? That's $15M for one of the best advertising campaigns in human history. It is baffling to me that they haven't done it.
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 things and 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!
By treating everything as an emergency, you're treating nothing as an emergency. Effectively, what I've learned from my time with Fable isn't the science and math I've wanted to do. It's how to step past your "safety" filters.
You are training millions of users one denied prompt at a time. And you can't tell who is truly malicious and who isn't, because they're all a threat. If everyone's a threat, then no one's a threat.
I am happy to pay for your service well into the future. Please just let me use it.
What does it matter which tool I use when I hit the limit?
And even if you pay-as-you-go Anthropic seems to prefer their own tools
https://news.ycombinator.com/item?id=48788599
A third party harness may have a misconfiguration of prompt caching, leading to more load on Anthropic's servers, they could also have wildly different usage patterns (Hermes, openclaw etc), also making it hard to predict the load. If you constrain everyone to use your own harness, at least you're in control of that.
The first point is sort of funny though, because their own harness had a misconfigured prompt caching for several months during the times where they were crying about third party harnesses (/resume was busting cache all the time).
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
Nowadays I went from Claude 20x to 5x and been using the GLM model on OpenCode... No regrets.
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.
If agi why not agi shaped?
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.
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.