They’re not literally blocking OpenCode. You can use OpenCode with their API like any other tool.
They’ve blocked the workaround OpenCode was using to access a private API that was metered differently.
Any tool that used that private endpoint would be blocked. They’re not pushing an agenda. They’re just enforcing their API terms like they would for any use.
after they exploited us by training without any limits on code without licensing it (including GPLed code) they now scramble to ban and restrict when we want to do the same to them. that's the schadenfreude...
They may however be obligated to not give customers access to their services at a discounted rate either - predatory pricing is at least some of the time and in some jurisdictions illegal.
Predatory pricing is selling something below cost to acquire/maintain market dominance.
The Claude subscription used for Claude Code is to all appearances being sold substantially below the cost to run it, and it certainly seems that this is being done to maintain Claude Code's market dominance and force out competitors who cannot afford to subsidize LLM inference in the same way such as OpenCode.
It's not a matter of there being a public API, I don't believe they are obligated to offer one at all, it's a matter of the Claude Subscription being priced fairly so that OpenCode (on top of, say, gemini) can be competitive.
> Predatory pricing is selling something below cost to acquire/maintain market dominance.
Yet they have to acquire market dominance in a meaningful market first if you want to prosecute, otherwise it's just a failed business strategy. Like that company selling movie tickets bellow cost.
The modern consumer benefit doctrine means predatory pricing is impossible to prosecute in 99% of cases. I’m not saying it’s right, but legally it is toothless.
This is true... in the US (though there is still that 1%). Anthropic operates globally and the US isn't the only country who ever realized it might be an issue.
The API is really expensive compared to a Max subscription! So they're probably making a lot of money (or at least losing much less) via the API. I don't think it's going anywhere. Worst case scenario they could raise the price even more.
The Claude subscription (i.e. the pro and max plans, not the API) is sold at what appears to be well below cost in what appears to be a blatant attempt to preserve/create market dominance for claude code, destroying competitors by making it impossible to compete without also having a war chest of money to give away.
You’re making a big assumption. LLM providers aren’t necessarily taking a loss on the marginal cost of inference. It’s when you include R&D and training costs that it requires the capital inputs. They’ve come out and said as much.
The Claude Code plans may not be operating at a loss either. Most people don’t use up 100% of their plan. Few people do. A lot of it goes idle.
Are you suggesting Anthropic has a “duty to deal” with anyone who is trying to build competitive products to Claude Code, beyond access to their priced API? I don’t think so. Especially not to a product that’s been breaking ToS.
A regulatory duty to deal is the opposite of setting your own terms. Yes, citing a ToS is acceptable in this scenario. We can throw ToS out if we all believed in duty to deal.
They cannot actually do this as long as they keep Claude code open source. It is always going to be trivial to replicate how it sends requests in a third party tool.
Yes, there is no source code in here. This is their scripts / tooling / prompts repo. The actual code that powers their CC terminal CLI does not exist anywhere on their public GitHub
It is available on npm but it’s a wasm file last I checked. You also don’t need it to find their endpoints, people are just seeing what networks calls are made when they use Claude Code and then try to get other agents to call those endpoints.
The hard part is that they have an Anthropic-compatible API that’s different than completion/responses.
Because your subscription depends on the very API business.
Anthropic's cogs is rent of buying x amount of h100s. cost of a marginal query for them is almost zero until the batch fills up and they need a new cluster. So, API clusters are usually built for peak load with low utilization (filled batch) at any given time. Given AI's peak demand is extremely spiky they end up with low utilization numbers for API support.
Your subscription is supposed to use that free capacity. Hence, the token costs are not that high, hence you could buy that. But it needs careful management that you dont overload the system. There is a claude code telemetry which identifies the request as lower priority than API (and probably decide on queueing + caching too). If your harness makes 10 parallel calls everytime you query, and not manage context as well as claude code, its overwhelming the system, degrading the performance for others too. And if everyone just wants to use subscription and you have no api takers, the price of subscription is not sustainable anyway. In a way you are relying on others' generosity for the cheap usage you get.
Its reasonable for a company to unilaterally decide how they monetize their extra capacity, and its not unjustified to care. You are not purchasing the promise of X tokens with a subscription purchase for that you need api.
> Your subscription is supposed to use that free capacity. Hence, the token costs are not that high, hence you could buy that. But it needs careful management that you dont overload the system. There is a claude code telemetry which identifies the request as lower priority than API (and probably decide on queueing + caching too). If your harness makes 10 parallel calls everytime you query, and not manage context as well as claude code, its overwhelming the system, degrading the performance for others too. And if everyone just wants to use subscription and you have no api takers, the price of subscription is not sustainable anyway. In a way you are relying on others' generosity for the cheap usage you get.
I understand what you mean but outright removing the ability for other agents to use the claude code subscription is still really harsh
If telemetry really is a reason (Note: I doubt it is, I think the marketing/lock-ins aspect might matter more but for the sake of discussion, lets assume so that telemetry is in fact the reason)
Then, they could've simply just worked with co-ordination with OpenCode or other agent providers. In fact this is what OpenAI is doing, they recently announced a partnership/collaboration with OpenCode and are actively embracing it in a way. I am sure that OpenCode and other agents could generate telemetry or atleast support such a feature if need be
From what i have read on twitter. People were purchasing max subs and using it as a substitute for API keys for their startups. Typical scrappy startup story but this has the same bursty nature as API in temrs of concurrency and parallel requests. They used the Opencode implementation. This is probably one of the triggers because it screws up everything.
Telemetry is a reason. And its also the mentioned reason. Marketing is a plausible thing and likely part of the reason too, but lock-in etc. would have meant this would have come way sooner than now. They would not even be offering an API in that case if they really want to lock people in. That is not consistent with other actions.
At the same time, the balance is delicate. if you get too many subs users and not enough API users, then suddenly the setup is not profitable anymore. Because there is less underused capacity available to direct subs users to. This probably explains a part of their stance too, and why they havent done it till now. Openai never allowed it, and now when they do, they will make more changes to the auth setup which claude did not. (This episode tells you how duct taped whole system was at ant. They used the auth key to generate a claude code token, and just used that to hit the API servers).
Demand-aggregation allows the aggregator to extract the majority of the value. ChatGPT the app has the biggest presence, and therefore model improvements in Claude will only take you so far. Everyone is terrified of that. Cursor et al. have already demonstrated to model providers that it is possible to become commoditized. Therefore, almost all providers are seeking to push themselves closer to the user.
This kind of thing is pretty standard. Nobody wants to be a vendor on someone else's platform. Anthropic would likely not complain too much about you using z.ai in Claude Code. They would prefer that. They would prefer you use gpt-5.2-high in Claude Code. They would prefer you use llama-4-maverick in Claude Code.
Because regardless of how profitable inference is, if you're not the closest to the user, you're going to lose sooner or later.
Because Claude Code is not a profitable business, it's a loss leader to get you to use the rest of their token inference business. If you were to pay for Claude Code by using the normal API, it would be at least 5x the cost, if not more.
he may not be entirely correct, but Claude Code plans are significantly better than the API plan, 100$ plan may not be as cost effective but for 18$ you can get like 5x usage of the API plan.
I've seen dozens of "experts" all over the internet claim that they're "subsidizing costs" with the coding plans despite no evidence whatsoever. Despite the fact that various sources from OpenAI, Deepseek, model inference providers have suggested the contrary, that inference is very profitable with very high margins.
Just looking at my own usage at work, we’re spending around $50/day on OpenAI API credits (with Codex). With Claude Code I get higher usage limits for $200/month, or around $8/day. Probably the equivalent from OpenAI is around $100/day of API credits.
Maybe OpenAI has a 12x markup on API credits, or Anthropic is much better at running inference, but my best guess is that Anthropic is selling at a large loss.
You can't be comparing OpenAI API with Anthropic subscription. The comparison here is OpenAI Codex subscription with Anthropic subscription. And when you do that, it turns out that the Codex limits are a lot higher for the same price. So then if Anthropic is selling at a large loss, OpenAI is selling at an even much bigger one.
How am I gonna give you exact price savings, when on $18 amount of work you can do it is variable, while $100 on API only goes a limited amount. You can exhaust $100 on API in one work day easily. On $18 plan the limit resets daily or 12hrs, so you can keep coming back. If API pricing is correct, which it looks like because all top models have similar costs, then it is to believe that monthly plans are subsidised.
And if inference is so profitable why is OpenAI losing 100B a year
That made me imagine -- in the future when AI is much more advanced, maybe I could just prompt it with say "something sentimental to make my wife cry." I mean, I still came up with the idea and ultimately it's the thought that counts right. What's the limit here? Is this some sort of human emotion exploit, or a legitimate bonding experience?
It’s rarely the thought that counts. It’s the committed effort. Presents aren’t just nice because they needed those socks. More importantly, they’re a signifier that you consider the person to be worth thinking about. You value them enough to spend time and effort thinking about them. Then you followed through. This is why we don’t just give people money as a present.
The effort that you put in is often what people like most about a gift. Don’t try too hard to hack around that.
I'm going to draw this example out to make it more realistic.
"Say something sentimental to make my wife cry" you prompt. The computer comes back:
Ok, tell me a few things about your wife. How did you meet? What are her favorite things? Tell me about some great moments in your relationship. Tell me about some difficult moments in your relationship.
Ok, tell me a few things about you. What do you love about your wife? What have you struggled with?
Ten minutes of this kind of conversation and I'll bet the LLM can generate a pretty good hallmark card. It might not make your wife cry but she'll recognize it as something personal and special.
Four hours of this kind of conversation and you might very well get some output that would make your wife cry. It might even make you cry.
The work is adding context. And getting people to add meaningful, soul-touching context is not easy - just ask any therapist.
1. Wives aren't a monolith. The prompt is underspecified, or else individual taste and preciousness is dead.
2. No matter how good the tech today is (or isn't) getting, the responses are very low temperature. The reason it takes a human 4 hours to write the poem is because that is time spent exploring profoundly new structures and effects. Compare this to AI which is purpose-built to hone in on local optima of medians and clichés wherever possible.
> I mean, I still came up with the idea and ultimately it's the thought that counts right. What's the limit here?
Sociologically, devoid of AI discussion, I imagine the limit is the extent to which the ideas expressed in the poem aren't outright fabrications (e.g. complimenting their eyes when really you couldn't care less). As well, it does not sit right with humans if you attempt to induce profound feelings in them by your own less-than-profound feelings; it's not "just the thought," it's also the effort that socially signals the profundity of the thought.
Usually they are. Most people are surprisingly similar and predictable, which is why basic manipulation tactics are so successful. Sure, you have 10% of people who truly are special, but the other 90% has a collective orgasm while listening to whatever is the hottest pop star.
> The reason it takes a human 4 hours to write the poem is because that is time spent exploring profoundly new structures and effects.
Most likely dude spent 4 hours doing exactly the same things that everyone else does when making their first song. It's not like within these 4 hours he discovered a truly new technique of writing lyrics. Each instance of human life that wants to write songs needs to go through exactly the same learning steps, while AI does it just once and then can endlessly apply the results.
> it's not "just the thought," it's also the effort that socially signals the profundity of the thought.
In close relationships yes. When dealing with those you less care about, it's the result that matters.
I think expended effort is what counts here for these types of interactions, and how much of that effort is tailored to the specific person.
I mean, we're almost always standing on the shoulders of other people, and we're almost always using tools. But if the output is fully mechanical and automatic without being tailored for the specific person, it's hard to see it as personal in any way.
I think nearly 100% of blog posts are run through an LLM now. The author was lazy and went with the default LLM "tone" and so the typical LLM grammar usage and turns of phrase are too readily apparent.
It's really disheartening as I read blogs to get a perspective on what another human thinks and how their thought process works. If I wanted to chat with a LLM, I can open a new tab.
It's just fatigue from seeing the same people and themes repeatedly, non-stop, for the last X months on the site. Eventually you'd expect some tired reactions.
> "Just Eat Less" is roughly the way to lose weight
Maybe the messaging should be "eat healthier"? How many obese people cook for themselves and eat exclusively from the outer aisles of the grocery (fruits, vegetables, seafood, meat, eggs, dairy)?
I could be wrong but I have to imagine the average obese person has a terrible diet. Portion control won't work at that point, you're already doomed to fail.
To be fair, most people have a terrible diet, it's just that some lucky individuals have the metabolism to overcome it. It seems like those people are increasingly the exception and a bad benchmark for how humans should eat.
Differences in metabolism are very seldom the real reason. The people who claim they have a "slow" or "fast" metabolism can't back that up with actual RMR test results. They're just bad at estimating many calories they actually consume. This can go both ways.
> For years, despite functional evidence and scientific hints accumulating, certain AI researchers continued to claim LLMs were stochastic parrots: probabilistic machines that would: 1. NOT have any representation about the meaning of the prompt. 2. NOT have any representation about what they were going to say. In 2025 finally almost everybody stopped saying so.
It's interesting that Terrence Tao just released his own blog post stating that they're best viewed as stochastic generators. True he's not an AI researcher, but it does sound like he's using AI frequently with some success.
"viewing the current generation of such tools primarily as a stochastic generator of sometimes clever - and often useful - thoughts and outputs may be a more productive perspective when trying to use them to solve difficult problems" [0].
I get the impression that folks who have a strong negative reaction to the phrase "stochastic parrot" tend to do so because they interpret it literally or analogously (revealed in their arguments against it), when it is most useful as a metaphor.
(And, in some cases, a desire to deny the people and perspectives from which the phrase originated.)
What happened recently is that all the serious AI researches that were in the stochastic parrot side changed point of view but, incredibly, people without a deep understanding on such matters, previously exposed to such arguments, are lagging behind and still repeat arguments that the people who popularized them would not repeat again.
Today there is no top AI scientist that will tell you LLMs are just stochastic parrots.
You seem to think the debate is settled, but that’s far from true. It’s oddly controlling to attempt to discredit any opposition to this viewpoint. There’s plenty of research supporting the stochastic view of these models, such as Apple’s “Illusion” papers. Tao is also a highly respected researcher, and has worked with these models at a very high level - his viewpoint has merit as well.
The stochastic parrot framing makes some assumptions, one of them being that LLMs generate from minimal input prompts, like "tell me about Transformers" or "draw a cute dog". But when input provides substantial entropy or novelty, the output will not look like any training data. And longer sessions with multiple rounds of messages also deviate OOD. The model is doing work outside its training distribution.
It's like saying pianos are not creative because they don't make music. Well, yes, you have to play the keys to hear the music, and transformers are no exception. You need to put in your unique magic input to get something new and useful.
Ok, I'll say it: it's for AI datacenters to train chat bots.
reply