I really hope someone from any of those companies (if possible all of them) would publish a very clear statement regarding the following question: If I build a commercial app that allows my users to connect using their OAuth token coming from their ChatGPT/Claude etc. account, do they allow me (and their users) to do this or not?
I totally understand that I should not reuse my own account to provide services to others, as direct API usage is the obvious choice here, but this is a different case.
I am currently developing something that would be the perfect fit for this OAuth based flow and I find it quite frustrating that in most cases I cannot find a clear answer to this question. I don't even know who I would be supposed to contact to get an answer or discuss this as an independent dev.
EDIT: Some answers to my comment have pointed out that the ToS of Anthropic were clear, I'm not saying they aren't if taken in a vacuum, yet in practice even after this being published some confusion remained online, in particular regarding wether OAuth token usage was still ok with the Agent SDK for personal usage. If it happens to be, that would lead to other questions I personally cannot find a clear answer to, hence my original statement. Also, I am very interested about the stance of other companies on this subject.
Maybe I am being overly cautious here but I want to be clear that this is just my personal opinion and me trying to understand what exactly is allowed or not. This is not some business or legal advice.
I don't see how they can get more clear about this, considering they have repeatedly answered it the exact same way.
Subscriptions are for first-party products (claude.com, mobile and desktop apps, Claude Code, editor extensions, Cowork).
Everything else must use API billing.
The biggest reason why this is confusing is the Claude Agent SDK[0] will use subscription/oauth credentials if present. The terms update implies that there's some use cases where that's ok and other use cases (commercial?) where using their SDK on a user's device violates terms.
Had the same question, comment below quotes their docs saying Agent SDK using oAuth token is also not allowed.
The SDK is Claude Code in a harnesss, so it works with your credentials the same way CC does.
But they're stating you can only use your subscription for your personal usage, not someone else's for their usage in your product.
I honestly think they're being short sighted not just giving a "3rd party quota" since they already show users like 4 quotas.
If the fear is 3rd party agents screwing up the math, just make it low enough for entry level usage. I suspect 3rd party token usage is bi-modal where some users just need enough to kick tires, but others are min-maxing for how mamy tokens they can burn as if that's its own reward
And at that point, you might as well use OpenRouter's PKCE and give users the option to use other models..
These kinds of business decisions show how these $200.00 subscriptions for their slot/infinite jest machines basically light that $200.00 on fire, and in general how unsustainable these business models are.
Can't wait for it all to fail, they'll eventually try to get as many people to pay per token as possible, while somehow getting people to use their verbose antigentic tools that are able to inflate revenue through inefficient context/ouput shenanigans.
I think the subscription pricing exists because it’s a far more palatable way to bill people for day to day personal use.
I used Claude back when API per token pricing was the only option and it was bad for all the usual reasons pay-per-use sucks compared to flat billing: you’re constantly thinking about cost. Like trying to watch a Netflix video with a ticker in the corner counting up the cents you owe them.
I don’t understand your claim that they want people paying per token - the subscription is the opposite of that, and it also has upsides for them as a business since most people don’t saturate the usage limits, and the business gets to stuff a bunch of value-adds on a bundle offering which is generally a more lucrative and enticing consumer pricing model.
You are talking about Anthropic and indeed compared to OpenAI or GitHub Copilot they have seemed to be the ones with what I would personally describe as a more restrictive approach.
On the other hand OpenAI and GitHub Copilot have, as far as I know, explicitly allowed their users to connect to at least some third party tools and use their quotas from there, notably to OpenCode.
What is unclear to me is whether they are considering also allowing commercial apps to do that. For instance if I publish a subscription based app and my users pay for the app itself rather than for LLM inference, would that be allowed?
What if you wrap the service using their Agent SDK?
Quick question but what if I use claude code itself for the purpose?
https://github.com/rivet-dev/sandbox-agent/tree/main/gigacod... [I saw this inShow HN: Gigacode – Use OpenCode's UI with Claude Code/Codex/Amp] (https://news.ycombinator.com/item?id=46912682)
This can make Opencode work with Claude code and the added benefit of this is that Opencode has a Typescript SDK to automate and the back of this is still running claude code so technically should work even with the new TOS?
So in the case of the OP. Maybe Opencode TS SDK <-> claude code (using this tool or any other like this) <-> It uses the oauth sign in option of Claude code users?
Also, zed can use the ACP protocol itself as well to make claude code work iirc. So is using zed with CC still allowed?
> I don't see how they can get more clear about this, considering they have repeatedly answered it the exact same way.
This is confusing quite frankly, there's also the claude agent sdk thing which firloop and others talked about too. Some say its allowed or not. Its all confusing quite frankly.
That’s very clearly a no, I don’t understand why so many people think this is unclear.
You can’t use Claude OAuth tokens for anything. Any solution that exists worked because it pretended/spoofed to be Claude Code. Same for Gemini (Gemini CLI, Antigravity)
Codex is the only one that got official blessing to be used in OpenClaw and OpenCode, and even that was against the ToS before they changed their stance on it.
Is Codex ok with any other third party applications, or just those?
By default, assume no. The lack of any official integration guide should be a clear sign. Even saying that you reverse-engineer Codex for apps to pretend to be Codex makes it clear that this is not an officially endorsed thing to do
Interested to know this too
But why does it matter which program consumes the tokens?
They'll own entire pipeline interface, conduit, backend. Interface is what people get habitual to. If I am a regular user of Claude Code, I may not shift to competitor for 10-20% gains in cost.
They must be getting something out of it, because we sure aren't.
Cory Doctorow has a word for this..
They think their position is strong enough to lock users in. I'm not so sure.
It's enshittification - for those who didn't know.
They want that sweet vendor lock-in.
Presumably because their flat rate pricing is based off their ability to manage token use via their first-party tools.
A third-party tool may be less efficient in saving costs (I have heard many of them don't hit Anthropic LLMs' caches as well).
Would you be willing to pay more for your plan, to subsidize the use of third-party tools by others?
---
Note, afaik, Anthropic hasn't come out and said this is the reason, but it fits.
Or, it could also just be that the LLM companies view their agent tools as the real moat, since the models themselves aren't.
I think you're just trying to see ambiguity where it doesn't exist because the looser interpretation is beneficial to you. It totally makes sense why you'd want that outcome and I'm not faulting you for it. It's just that, from a POV of someone without stake in the game, the answer seems quite clear.
It is pretty obviously no. API keys billed by the token, yes, Oauth to the flat rate plans no.
> OAuth authentication (used with Free, Pro, and Max plans) is intended exclusively for Claude Code and Claude.ai. Using OAuth tokens obtained through Claude Free, Pro, or Max accounts in any other product, tool, or service — including the Agent SDK — is not permitted and constitutes a violation of the Consumer Terms of Service.
If you look at this tweet [1] and in particular responses under it, it still seems to me like some parts of it need additional clarification. For instance, I have seen some people interpret the tweet as meaning using the OAuth token is actually ok for personal experimentation with the Agent SDK, which can be seen as a slight contradiction with what you quoted. A parent tweet also mentioned the docs clean up causing some confusion.
None of this is legal advice, I'm just trying to understand what exactly is allowed or not.
Read the actual ToS. What you describe is NOT allowed.
That tweet is from a product leader on Claude Code itself...
A tweet is not a ToS.
Then they should speak to legal about fixing the ToS before making public statements about their intentions with it. It won't look good to show up at arbitration and have to explain why your public comms contradict your ToS.
What flatrate?
Pro and Max are both limited
Flat rate does not imply unlimited.
> OAuth authentication (used with Free, Pro, and Max plans) is intended exclusively for Claude Code and Claude.ai.
I think this is pretty clear - No.
Does https://happy.engineering/ need to use the API keys or can use oauth? It's basically a frontend for claude-cli.
It doesn't even touch auth right?
""" Usage policy
Acceptable use Claude Code usage is subject to the Anthropic Usage Policy. Advertised usage limits for Pro and Max plans assume ordinary, individual usage of Claude Code and the Agent SDK """
That tool clearly falls under ordinary individual use of Claude code. https://yepanywhere.com/ is another such tool. Perfectly ordinary individual usage.
https://yepanywhere.com/sdk-auth-clarification.html
The TOS are confusing because just below that section it talks about authentication/credential use. If an app starts reading api keys / credentials, that starts falling into territory where they want a hard line no.
If it's a wrapper that invokes the `claude` binary then I believe it's fine.
Is there a way to legally or even practically prevent this? `claude` CLI execution in a shell is certainly included in the subscription - it’s the product.
Anthropic has published a very clear statement. It's "no".
Usually, it is already stated in their documentation (auth section). If a statement is vague, treat it as a no. It is not worth the risk when they can ban you at any time. For example, ChatGPT allows it, but Claude and Gemini do not.
Maybe I am missing something from the docs of your link, but I unfortunately don't think it actually states anything regarding allowing users to connect and use their Codex quota in third party apps.
https://x.com/thdxr/status/2013010664776683644
I can't find anything official from OpenAI, but they have worked with the OpenCode people to support using your ChatGPT subscription in OpenCode.
From TFA: “OAuth authentication (used with Free, Pro, and Max plans) is intended exclusively for Claude Code and Claude.ai. Using OAuth tokens obtained through Claude Free, Pro, or Max accounts in any other product, tool, or service — including the Agent SDK — is not permitted and constitutes a violation of the Consumer Terms of Service.”
The comment you are responding to is about ChatGPT/Codex, not Claude.
They're not asking if Claude forbids it. They're asking if OpenAI (Codex, specifically) allows it.
One set of applications to build with subscription is to use the claude-go binary directly. Humanlayer/Codelayer projects on GitHub do this. Granted those are not ideal for building a subscription based business to use oathu tokens from Claude and OpenaAI. But you can build a business by building a development env and gating other features behind paywall or just offering enterprise service for certain features like vertical AI(redpanada) offerings knowledge workers, voice based interaction(there was a YC startup here the other day doing this I think), structured outputs and workflows. There is lots to build on.
I don't think it's a secret that AI companies are losing a ton of money on subscription plans. Hence the stricter rate limits, new $200+ plans, push towards advertising etc. The real money is in per-token billing via the API (and large companies having enough AI FOMO that they blindly pay the enormous invoices every month).
They are not losing money on subscription plans. Inference is very cheap - just a few dollars per million tokens. What they’re trying to do is bundle R&D costs with inference so they can fund the training of the next generation of models.
Banning third-party tools has nothing to do with rate limits. They’re trying to position themselves as the Apple of AI companies -a walled garden. They may soon discover that screwing developers is not a good strategy.
They are not 10× better than Codex; on the contrary, in my opinion Codex produces much better code. Even Kimi K2.5 is a very capable model I find on par with Sonnet at least, very close to Opus. Forcing people to use ONLY a broken Claude Code UX with a subscription only ensures they loose advantage they had.
> "just a few dollars per million tokens"
Google AI Pro is like $15/month for practically unlimited Pro requests, each of which take million tokens of context (and then also perform thinking, free Google search for grounding, inline image generation if needed). This includes Gemini CLI, Gemini Code Assist (VS Code), the main chatbot, and a bunch of other vibe-coding projects which have their own rate limits or no rate limits at all.
It's crazy to think this is sustainable. It'll be like Xbox Game Pass - start at £5/month to hook people in and before you know it it's £20/month and has nowhere near as many games.
I wonder how many people have a subscription and don’t fully utilize it. That’s free money for them, too.
> They are not losing money on subscription plans. Inference is very cheap - just a few dollars per million tokens. What they’re trying to do is bundle R&D costs with inference so they can fund the training of the next generation of models.
You've described every R&D company ever.
"Synthesizing drugs is cheap - just a few dollars per million pills. They're trying to bundle pharmaceutical research costs... etc."
There's plenty of legit criticisms of this business model and Anthropic, but pointing out that R&D companies sink money into research and then charge more than the marginal cost for the final product, isn't one of them.
I’m not saying charging above marginal cost to fund R&D is weird. That’s how every R&D company works.
My point was simpler: they’re almost certainly not losing money on subscriptions because of inference. Inference is relatively cheap. And of course the big cost is training and ongoing R&D.
The real issue is the market they’re in. They’re competing with companies like Kimi and DeepSeek that also spend heavily on R&D but release strong models openly. That means anyone can run inference and customers can use it without paying for bundled research costs.
Training frontier models takes months, costs billions, and the model is outdated in six months. I just don’t see how a closed, subscription-only model reliably covers that in the long run, especially if you’re tightening ecosystem access at the same time.
The secret is there is no path on making that back.
My crude metaphor to explain to my family is gasoline has just been invented and we're all being lent Bentley's to get us addicted to driving everywhere. Eventually we won't be given free Bentley's, and someone is going to be holding the bag when the infinite money machine finally has a hiccup. The tech giants are hoping their gasoline is the one that we all crave when we're left depending on driving everywhere and the costs go soaring.
Why? Computers and anything computer related have historically been dropping in prices like crazy year after year (with only very occasional hiccups). What makes you think this will stop now?
Commodity hardware and software will continue to drop in price.
Enterprise products with sufficient market share and "stickiness", will not.
For historical precedent, see the commercial practices of Oracle, Microsoft, Vmware, Salesforce, at the height of their power.
> Commodity hardware and software will continue to drop in price.
The software is free (citation: Cuda, nvcc, llvm, olama/llama cpp, linux, etc)
The hardware is *not* getting cheaper (unless we're talking a 5+ year time) as most manufacturers are signaling the current shortages will continue ~24 months.
Cars have also been dropping in price.
Evidence for this claim?
A few generations ago almost nobody could afford a car, now many low income families afford two.
Recent price trends for DRAM, SSDs, hard drives?
Short term squeeze, because building capacity takes time and real funding. The component manufacturers have been here before. Booms rarely last long enough to justify a build-out. If AI demand turns out to be sustained, the market will eventually adapt by building supply, and prices will drop. If AI demand turns out to be transient, demand will drop, and prices will drop.
In the GP's analogy, the Bentley can be rented for $3/day, but if you want to purchase it outright, it will cost you $3,000,000.
Despite the high price, the Bentley factory is running 24/7 and still behind schedule due to orders placed by the rental-car company, who has nearly-infinite money.
I like this analogy.
I also think we're, as ICs, being given Bentleys meanwhile they're trying to invent Waymos to put us all out of work.
Humans are the cost center in their world model.
how do I understand what is the sustainable pricing?
Depends on how you do the accounting. Are you counting inference costs or are you amortizing next gen model dev costs. "Inference is profitable" is oft repeated and rarely challenged. Most subscription users are low intensity users after all.
The question I have: how much are they _also_ losing on per-token billing?
From what I understand, they make money per-token billing. Not enough for how much it costs to train, not accounting for marketing, subscription services, and research for new models, but if they are used, they lose less money.
Finance 101 tldr explanation: The contribution margin (= price per token -variable cost per token ) this is positive
Profit (= contribution margin x cuantity- fix cost)
I agree; unfortunately when I brought up that they're losing before I get jumped on demanding me to "prove it" and I guess pointing at their balance sheets isn't good enough.
Why do you think they're losing money on subscriptions?
Because they're not saying they are making a profit
That doesn’t mean that the subscription itself is losing money. The margin on the subscription could be fine, but by using that margin to R&D the next model, the org may still be intentionally unprofitable. It’s their investment/growth strategy, not an indictment of their pricing strategy.
But why does it matter which program you use to consume the tokens?
The sounds like a confession that claude code is somewhat wasteful at token use.
No, it's a confession they have no moat other than trying to hold onto the best model for a given use case.
I find that competitive edge unlikely to last meaningfully in the long term, but this is still a contrarian view.
More recently, people have started to wise up to the view that the value is in the application layer
https://www.iconiqcapital.com/growth/reports/2026-state-of-a...
Honestly I think I am already sold on AI, who is the first company that is going to show us all how much it really costs and start enshitification? First to market wins right?
Not according to this guy who works on Claude Code: https://x.com/trq212/status/2024212378402095389?s=20
What a PR nightmare, on top of an already bad week. I’ve seen 20+ people on X complaining about this and the related confusion.
No, it is prohibited. They're just updating the docs to be more clear about their position, which haven't changed. Their docs was unclear about it.
Yes, it was always prohibited, hence the OpenCode situation one or two months ago.
woof, does Anthropic not have a comms team and a clear comms policy for employees that aren’t on that comms team?
Incorrect, the third-party usage was already blocked (banned) but it wasn't officially communicated or documented. This post is simply identifying that official communication rather than the inference of actual functionality.
At this point, where Kimi K2.5 on Bedrock with a simple open source harness like pi is almost as good the big labs will soon have to compete for users,... openai seems to know that already? While anthropic bans bans bans
The economic tension here is pretty clear: flat-rate subscriptions are loss leaders designed to hook developers into the ecosystem. Once third parties can piggyback on that flat rate, you get arbitrage - someone builds a wrapper that burns through $200/month worth of inference for $20/month of subscription cost, and Anthropic eats the difference.
What is interesting is that OpenAI and GitHub seem to be taking the opposite approach with Copilot/OpenCode, essentially treating third-party tool access as a feature that increases subscription stickiness. Different bets on whether the LTV of a retained subscriber outweighs the marginal inference cost.
Would not be surprised if this converges eventually. Either Anthropic opens up once their margins improve, or OpenAI tightens once they realize the arbitrage is too expensive at scale.
these subscriptions have limits.. how could someone use $200 worth on $20/month.. is that not the issue with the limits they set on a $20 plan, and couldn't a claude code user use that same $200 worth on $20/month? (and how do i do this?)
The limits in the max subscriptions are more generous and power users are generating loss.
I'm rather certain, though cannot prove it, that buying the same tokens would cost at least 10x more if bought from API. Anecdotally, my cursor team usage was getting to around 700$ / month. After switching to claude code max, I have so far only once hit the 3h limit window on the 100$ sub.
What Im thinking is that Anthropic is making loss with users who use it a lot, but there are a lot of users who pay for max, but don't actually use it.
With the recent improvements and increase of popularity in projects like OpenClaw, the number of users that are generating loss has probably massively increased.
I'd agree on this. I ended up picking up a Claude Pro sub and am very less than impressed at the volume allowance. I generally get about a dozen queries (including simple follow up/refinements/corrections) across a relatively small codebase, with prompts structured to minimize the parts of the code touched - and moving onto fresh contexts fairly rapidly, before getting cut off for their ~5 hour window. Doing that ~twice a day ends up getting cut off on the weekly limit with about a day or two left on it.
I don't entirely mind, and am just considering it an even better work:life balance, but if this is $200 worth of queries, then all I can say is LOL.
The usage limit on your $20/month subscription is not $20 of API tokens (if it was, why subscribe?). Its much much higher, and you can hit the equivalent of $20 of API usage in a few days.
I think you've stated this in reverse.
API limits are infinite but you'd blow through $20 of usage in a maybe 1 hours or less of intense Opus use.
The subscription at $20/mo (or $200) allows for vastly more queries than $20 would buy you but you are constrained by hourly/weekly limits.
In practice, you're a _lot_ slower to complete a task on $20/mo vs api but you cap your costs.
So I’m not allowed to use the $20 plan and max out its limits?
Max out on their terms, not yours.
Their bet is that most people will not fill up 100% of their weekly usage for 4 consecutive weeks of their monthly plan, because they are humans and the limits impede long running tasks during working hours.
You can max it out via first party clients only.
I dont like it either, but its not an unreasonable restriction.
OK I hope someone from anthropic reads this. Your API billing makes it really hard to work with it in India. We've had to switch to openrouter because anthropic keeps rejecting all the cards we have tried. And these are major Indian banks. This has been going on for MONTHS
It’s the same here in Hong Kong. I can’t use any of my cards (personal or corporate) for OpenAI or Anthropic.
Have to do everything through Azure, which is a mess to even understand.
The pressure is to boost revenue by forcing more people to use the API to generate huge numbers of tokens they can charge more for. LLMs are becoming common commodities as open weight models keep catching up. There are similarities with pirating in the 90s when users realize they can ctrl+c ctrl+v to copy a file/model and you don't need to buy a cd/use their paid API.
And that is how it should be - the knowledge that the LLM trained on should be free, and cannot (and should never be) gatekept behind money.
It's merely the hardware that should be charged for - which ought to drop in price if/when the demand for it rises. However, this is a bottleneck at the moment, and hard to see how it gets resolved amidst the current US environment on sanctioning anyone who would try.
Is there no value in how the training was done such that it's accessible via inference in a particularly useful way?
That value is there, but google has decided to give it away as public knowledge (ala, their transformer paper).
And i would also argue that the researchers doing this are built on shoulders of other public knowledge - things funded by public institutions with taxpayer money.
Just a friendly reminder also to anyone outside the US that these subscriptions cannot be used for commercial work. Check the consumer ToS when you sign up. It’s quite clear.
AI is the new high-end gym membership. They want you to pay the big fee and then not use what you paid for. We'll see more and more roadblocks to usage as time goes on.
This feels more like the gym owner clarifying it doesn't want you using their 24-hour gym as a hotel just because you find their benches comfortable to lie down on, rather than a "roadblock to usage"
Thariq has clarified that there are no changes to how SDK and max suscriptions work:
https://x.com/i/status/2024212378402095389
---
On a different note, it's surprising that a company that size has to clarify something as important as ToS via X
> On a different note, it's surprising that a company that size has to clarify something as important as ToS via X
Countries clarify nation policy on X. Seriously it feels like half of the EU parliament live on twitter.
FYI a Twitter post that contradicts the ToS is NOT a clarification.
What's wrong with using X?
In the case you are asking in good faith, a) X requires logging in to view most of its content, which means that much of your audience will not see the news because b) much of your audience is not on X, either due to not having social media or have stopped using X due to its degradation to put it generally.
I'm not signed in but I can view the above linked tweet just fine.
Plus it's not a real clarification in anyway. It's just PR. Even if it's posted on Mastodon or Github or anywhere, I highly doubt you can use it to defend yourself if you get banned from violating their ToS.
You can’t view answers and the tweet threat. You need to know every single tweet. You can’t open the politician‘s feed so you have to know that there is a tweet and which it is to get information.
c) it is controlled by a direct competitor and can bury / promote your customer communication at will.
Elon has enough sense not to cross that particular bridge.
He was quick to ban links to Mastodon when it was on the rise, I'm not sure why he'd treat SpaceX/xAI competitors any differently.
Not bad per se but how much legal weight does it actually carry?
I presume zero.. but nonetheless seems like people will take it as valid anyway.
That can be dangerous I think.
Going to keep using the agents sdk with my pro subscription until I get banned. It's not openclaw it's my own project. It started by just proxying requests to claude code though the command line, the sdk just made it easier. Not sure what difference it makes to them if I have a cron job to send Claude code requests or an agent sdk request. Maybe if it's just me and my toy they don't care. We'll see how the clarify tomorrow.
From the legal docs:
> Authentication and credential use
> Claude Code authenticates with Anthropic’s servers using OAuth tokens or API keys. These authentication methods serve different purposes:
> OAuth authentication (used with Free, Pro, and Max plans) is intended exclusively for Claude Code and Claude.ai. Using OAuth tokens obtained through Claude Free, Pro, or Max accounts in any other product, tool, or service — including the Agent SDK — is not permitted and constitutes a violation of the Consumer Terms of Service.
> Developers building products or services that interact with Claude’s capabilities, including those using the Agent SDK, should use API key authentication through Claude Console or a supported cloud provider. Anthropic does not permit third-party developers to offer Claude.ai login or to route requests through Free, Pro, or Max plan credentials on behalf of their users.
> Anthropic reserves the right to take measures to enforce these restrictions and may do so without prior notice.
why wouldn't they just make it so the SDK can't use claude subs? like what are they doing here?
When your company happens upon a cash cow, you can either become a milk company or a meat company.
That page is... confusing.
> Advertised usage limits for Pro and Max plans assume ordinary, individual usage of Claude Code and the Agent SDK.
This is literally the last sentence of the paragraph before the "Authentication and credential use"
Their model actually doesn't have that much of a moat if at all. Their agent harness also doesn't, at least not for long. Writing an agent harness isn't that difficult. They are desperately trying to stay in power. I don´t like being a customer of this company and am investing lots of my time in moving away from them completely.
They are obviously losing money on these plans, just like all of the other companies in the space.
They are all desperately trying to stay in power, and this policy change (or clarification) is a fart in the wind in the grand scheme of what's going on in this industry.
there’s a million small scale AI apps that just aren’t worth building because there’s no way to do the billing that makes sense. If anthropic wanted to own that market, they could introduce a bring-your-own-Claude metaphor, where you login with Claude and token costs get billed to your personal account (after some reasonable monthly freebies from your subscription).
But the big guys don’t seem interested in this, maybe some lesser known model will carve out this space
This is going to happen. Unfortunately.
I shudder to think what the industry will look like if software development and delivery becomes like Youtubing, where the whole stack and monetization is funneled through a single company (or a couple) get to decide who gets how much money.
I am a bit worried that this is the situation I am in with my (unpublished) commercial app right now: one of the major pain points I have is that while I have no doubt the app provides value in itself, I am worried about how many potential users will actually accept paying inference per token...
As an independent dev I also unfortunately don't have investors backing me to subsidize inference for my subscription plan.
I recommend kimi. It's possible for people to haggle with it to get cheap for the first month and as such try out your project and best part of the matter is that kimi intentionally supports api usage in any of their subscribed plan and they also recently changed their billing to be more token usage based like others instead of their previous tool calling limits
It's seriously one of the best models. very comparable to sonnet/opus although kimi isn't the best in coding. I think its a really great solid model overall and might just be worth it in your use case?
Is the use case extremely coding intensive related (where even some minor improvement can matter for 10-100x cost) or just in general. Because if not, then I can recommend Kimi.
>> there’s a million small scale AI apps that just aren’t worth building because there’s no way to do the billing that makes sense
Maybe they are not worth building at all then. Like MoviePass wasn’t.
In enterprise software, this is an embedded/OEM use case.
And historically, embedded/OEM use cases always have different pricing models for a variety of reasons why.
How is this any different than this long established practice?
It's not, but do you really think the people having Claude build wrappers around Claude were ever aware of how services like this are typically offered.
I got banned for violating terms of use apparently, but I'm mystified as to what I rule I broke, and appealing just vanishes into the ether.
I think this is shortsighted.
The markets value recurring subscription revenue at something like 10x “one-off” revenue, Anthropic is leaving a lot of enterprise value on the table with this approach.
In practice this approach forces AI apps to pay Anthropic for tokens, and then bill their customers a subscription. Customers could bring their own API key but it’s sketchy to put that into every app you want to try, and consumers aren’t going to use developer tools. And many categories of free app are simply excluded, which could in aggregate drive a lot more demand for subscriptions.
If Anthropic is worried about quota, seems they could set lower caps for third-party subscription usage? Still better than forcing API keys.
(Maybe this is purely about displacing other IDE products, rather than a broader market play.)
I think they are smart making a distinction between a D2C subscription which they control the interface to and eat the losses for vs B2B use where they pay for what they use.
Allows them to optimize their clients and use private APIs for exclusive features etc. and there’s really no reason to bootstrap other wannabe AI companies who just stick a facade experience in front of Anthropic’s paying customer.
> eat the losses
Look at your token usage of the last 30 days in one of the JSON files generated by Claude Code. Compare that against API costs for Opus. Tell me if they are eating losses or not. I'm not making a point, actually do it and let me know. I was at 1 million. I'm paying 90 EUR/m. That means I'm subsidizing them (paying 3-4 times what it would cost with the API)! And I feel like I'm a pretty heavy user. Although people running it in a loop or using Gas Town will be using much more.
There's no decision to be made here, it's just way too expensive to have 3rd parties soak up the excess tokens, that's not the product being sold.
Especially as they are subsidized.
That’s not true, the market loves pay per use, see ”cloud”. It outperforms subscriptions by a lot, it’s not ”one-off”. And your example is not how companies building on top tend to charge, you either have your own infrastructure (key) or get charged at-cost + fees and service costs.
I don’t think Anthropic has any desire to be some B2C platform, they want high paying reliable customers (B2B, Enterprise).
> the market loves pay per use, see ”cloud”.
Cloud goes on the books as recurring revenue, not one-off; even though it's in principle elastic, in practice if I pay for a VM today I'll usually pay for one tomorrow.
(I don't have the numbers but the vast majority of cloud revenue is also going to be pre-committed long-term contracts from enterprises.)
> I don’t think Anthropic has any desire to be some B2C platform
This is the best line of argument I can see. But still not clear to me why my OP doesn't apply for enterprise, too.
Maybe the play is just to force other companies to become MCPs, instead of enabling them to have a direct customer relationship.
Is this a direct shot at things like OpenClaw, or am I reading it wrong?
They even block Claude Code of you've modified it via tweakcc. When they blocked OpenCode, I ported a feature I wanted to Claude Code so I could continue using that feature. After a couple days, they started blocking it with the same message that OpenCode gets. I'm going to go down to the $20 plan and shift most of my work to OpenAI/ChatGPT because of this. The harness features matter more to me than model differences in the current generation.
Opencode as well. Folks have been getting banned for abusing the OAuth login method to get around paying for API tokens or whatever. Anthropic seems to prefer people pay them.
its not that innocent.
a 200 dollar a month customer isn't trying to get around paying for tokens, theyre trying to use the tooling they prefer. opencode is better in a lot of ways.
tokens get counted and put against usage limits anyway, unless theyre trying to eat analytics that are CC exclusive they should allow paying customers to consume to the usage limits in however way they want to use the models.
Anthropic is offering a steep discount in their plans. I highly doubt they want you using it in a harness where you can trivially switch away when someone else releases a better model
> they should allow paying customers to consume to the usage limits in however way they want to use the models.
I think I agree, but it's their business to run however they like. They have competition if we don't like it.
A $200/m max subscriber using OpenCode and not wanting to use API keys with pay-per-token pricing is very clearly trying to get around paying for tokens.
Is there any limits to that users 200/month? Why should they not be able to use the limits to the extent from other tools?
If openclaw chews my 200/month up in 15 days... I don't get more requests for free
For sure, yes. They already added attempts to block opencode, etc.
Not surprised, its the official stance by Anthropic.
I'm more surprised by people using subscription auth for OpenClaw when its officially not allowed.
What is the point of developing against the Agent SDK after this change.
At this point, are there decent alternatives to Anthropic models for coding that allow third-party usage?
OpenAI have been very generous in their plans in terms of token and what you use it with. Is Codex better or as good as Opus for coding? No. Is it a decent alternative? Very.
Kimi is amazing for this. They offer API usage as well iirc if you buy their subscription.
Also the .99c deal has API Access
Yes!
Why does it matter to Anthropic if my $200 plan usage is coming from Claude Code or a third party?
Doesn’t both count towards my usage limits the same?
If you buy a 'Season's Pass' for Disneyland, you cant 'sublet' it to another kid to use on the days you don't; It's not really buying a 'daily access rate'.
Anthropic subs are not 'bulk tokens'.
It's not an unreasonable policy and it's entirely inevitable that they have to restrict.
I’m not subletting my sub to anyone. I’m the only one using the third party harness.
I’m using their own SDK in my own CLI tool.
It’s not a literal sublet to someone else, it’s subletting your tokens to another tool.
At its core it’s a tragedy of commons situation. Using a third party tool like OpenClaw is augmenting your usage far beyond what was anticipated when the subscription plan was made.
Same deal for unlimited storage on drive until people started abusing it.
My Claude sub isn’t unlimited.
I didn’t set the limits on the plan; change those if it’s a problem, not irritate your customer base.
You have strong dedication towards taking things literally.
The issue is not that it's limited or unlimited, but rather about expected token usage across a user cohort. When you set a usage limit on something like Claude, or a gym, or a tutoring center, you need to do two things at once; set the limit high enough to attract the aspirations of your intended client base ("oh good this gym lets me go every day of the month if I want to"), but priced accurately enough so that you actually turn a profit on the average usage across most users (you ended up going 20 times the first month, but settled into 15 times a month after).
If there was suddenly a drug that you could take that would, while you slept, make your body walk to the gym and workout, so that you could max out that usage, the gym would be entitled to adjust either the pricing, the limit, or prohibit going to the gym while on the drug, given that they can't actually sustain all their members going every day.
As a correction, I've done some reading and when I said tragedy of the commons, what would fit better is a "congestion externality in a club good".
It’s still me going to Disneyland, I just take a different route
Disingenuous analogy.
It's more buying a season pass for Disneyland, then getting told you can't park for free if you're entering the park even though free parking is included with the pass. Still not unreasonable, but brings to light the intention of the tool is to force the user into an ecosystem rather.
It's not a disingenuous analogy ... whatever it is.
But 'you can't park even though the ticket includes parking' is not an appropriate analogy because 3rd party use is definitely not intended. They did not 'state one thing' and the 'disallow it'.
This is a pretty straight forward case of people using their subscription for 'adjacent' use, and Anthropic being more explicit about it.
There's nothing fancy going on here.
Any user who is using a third-party client is likely self-selected into being a power user who is less profitable.
They don't get as much visibility into your data, just the actual call to/from the api. There's so much more value to them in that, since you're basically running the reinforcement learning training for them.
Probably because the $20 plan is essentially a paid demo for the higher plans.
Increasing the friction of switching providers as much as possible is part of their strategy to push users to higher subscription tiers and deny even scraps to their competitors.
They're losing money on this $200 plan and they're essentially paying you to make you dependent on Claude Code so they can exploit this (somehow) in the future.
It's a bizarre plan because nobody is 'dependent' on Claude Code; we're begging to use alternatives. It's the model we want!
You’re not really paying for the model, you’re paying for the tool, the ecosystem, and the application layer around it.
Sonnet 4.6 in CC doesn’t behave the same way as Sonnet 4.6 in Antigravity.
When using Claude Code, it's possible to opt out of having one's sessions be used for training. But is that opt out for everything? Or only message content, such that there could remain sufficient metadata to derive useful insight from?
how can they even enforce this? can't you just spoof all your network requests to appear like it's coming from claude code?
in any case Codex is a better SOTA anyways and they let you do this. and if you aren't interested in the best models, Mistral lets you use both Vibe and their API through your vibe subscription api key which is incredible.
> how can they even enforce this?
Many ways, and they’re under no obligation to play fair and tell you which way they’re using at any given time. They’ve said what the rules are, they’ve said they’ll ban you if they catch you.
So let’s say they enforce it by adding an extra nonstandard challenge-response handshake at the beginning of the exchange, which generates a token which they’ll expect on all requests going forward. You decompile the minified JS code, figure out the protocol, try it from your own code but accidentally mess up a small detail (you didn’t realize the nonce has a special suffix). Detected. Banned.
You’ll need a new credit card to open a new account and try again. Better get the protocol right on the first try this time, because debugging is going to get expensive.
Let’s say you get frustrated and post on Twitter about what you know so far. If you share info, they’ll probably see it eventually and change their method. They’ll probably change it once a month anyway and see who they catch that way (and presumably add a minimum Claude Code version needed to reach their servers).
They’ve got hundreds of super smart coders and one of the most powerful AI models, they can do this all day.
the internet has hundreds of thousands of super smart coders with the most powerful ai models as well, I think it's a bit harder than you're assuming.
you just need to inspect the network traffic with Claude code and mimic that
When they blocked OpenCode, I was in the middle of adding a feature. I don't think it's possible to mimic CC in an undetectable way and have the feature work.
The feature allows the LLM to edit the context. For example, you can "compact" just portions of the conversation and replace it with a summary. Anthropic can see that the conversation suddenly doesn't share the same history as previous API calls.
In fact, I ported the feature to Claude Code using tweakcc, so it literally _is_ Claude Code. After a couple days they started blocking that with the same message that they send when they block third party tools.
OK sure, but you’d better hope Claude Code gets it right on the first try, or that’s $200 down the drain. Also, what if the detection mechanism involves a challenge-response that happens once a week? Or randomly a couple times a month? Or after 15 minutes of use from a new public IP? Or arbitrarily if you ask it to code something with a particular topic?
There are lots of ways they could be doing this. And remember again, if they get you, they don’t have to tell you how they got you (so you might not be able to even glean information in return for the $200 you’d be losing).
Sure the internet has hundreds of thousands of super smart coders, but the subset who are willing to throw money and credit cards down the drain in order to maintain a circumvention strategy for something like this is pretty low. I’m sure a few people will figure it out, but they won’t want to tell anyone lest Anthropic nerf their workaround, so I doubt that exploits of this will become widespread.
And if you’re Anthropic, that’s probably good enough.
Then just run Claude Code in a PTY and proxy requests in and out?
> Then just run Claude Code in a PTY and proxy requests in and out?
Exactly something I said too. There are projects which can do this and hook natively to opencode and even its sdk/api.
https://news.ycombinator.com/item?id=47069299#47070204 (I list a project which does this)
I really don't know how anthropic can somehow detect something like this.
If your service's traffic is literally indistinguishable from Claude Code, then all it can do is what Claude Code does. Then why are the users going to choose your service instead of Claude Code?
Maybe because the UI doesn't flicker? There's a lot you can do to improve the UI for starters, and then the harness around the agent could also be improved upon, as long as the prompts are the same.
easily "bypassable", trust me :)
see my comment here but I think instead of worrying about the decompile minified JS code etc., you can just essentially use claude code in the background and still do it even using opencode/its SDK thus giving sort of API access over CC subscription https://news.ycombinator.com/item?id=47069299#47070204
I am not sure how they can detect this. I can be wrong, I usually am but I think its still possible to use CC etc. even after this change if you really wanted to
But at this point, to me the question of GP that is that is it even worth it is definitely what I am thinking?
I think not. There are better options out there, they mentioned mistral and codex and I think kimi also supports maybe GLM/z.ai as well
Pretty easy to enforce it - rather than make raw queries to the LLM Claude Code can proxy through Anthropic's servers. The server can then enforce query patterns, system prompts and other stuff that outside apps cannot override.
And once all the Claude subscribers move over to Codex subscriptions, I'd bet a large sum that OpenAI will make their own ToS update preventing automated/scripted usage.
They can't catch everything but they can make your product you're building on top of it non viable when it gets popular enough to look for, like they did with opencode.
at least with open code you can just use a third party plugin to authenticate
> how can they even enforce this?
I would think that different tools would probably have different templates for their prompts?
You could tell by the prompt being used.
We don’t enforce speed limits, but it sucks when you get caught.
OpenAI will adjust, their investors will not allow money to be lost on ”being nice” forever, not until they’re handsomely paid back at least.
Their moat is evaporating before our eyes. Anthropic is Microsoft's side piece, but Microsoft is married with kids to OpenAI.
And OpenAI just told Microsoft why they shouldn't be seeing Anthropic anymore; Gpt-5.3-codex.
RIP Anthropic.
OpenAI has endorsed OAuth from 3rd party harnesses, and their limits are way higher. Use better tools (OpenCode, pi) with an arguably better model (xhigh reasoning) for longer …
I am looking forward to switching to OpenAI once my claude max account is banned for using pi....
My alt Google accounts were all banned from Gemini access. Luckily Google left my main account alone. They are all cracking down.
From 3rd party AI app use?
Codex has now caught up to Claude Opus and this is a defensive move by Anthropic
How does this impact open router?
Can’t this restriction for the time being be bypassed via -p command line flag?
OpenRouter uses the API and does not use any subscription auth.
So here goes my OpenClaw integration with Anthropic via OAuth… While I see their business risk I also see the onboarding path for new paying customers. I just upgraded to Max and would even consider the API if cost were controllable. I hope that Anthropic finds a smart way to communicate with customers in a constructive way and offers advice for the not so skilled OpenClaw homelabbers instead of terminating their accounts… Is anybody here from Anthropic that could pick up that message before a PR nightmare happens?
Sonnet literally just recommended using a subscription token for openclaw. Even anthropic's own AI doesn't understand its own TOS.
Sonnet was not trained with this information and extremely-recent-information-without-access-to-a-Web-Search-tool is the core case of hallucination.
Sonnet does have search available FYI.
LLMs are lazy about using Search unless explicitly prompted to do so (either explicitly or by implying it with phrases such as "recent").
In the OP's case, there is no motivation for the LLM to perform a Search.
Oh crap. I just logged into HN to ask if anyone knew of a working alternative to the Claude Code client. It's lost Claude's work multiple times in the last few days, and I'm ready to switch to a different provider. (4.6 is mildly better than 4.5, but the TUI is a deal breaker.)
So, I guess it's time to look into OpenAI Codex. Any other viable options? I have a 128GB iGPU, so maybe a local model would work for some tasks?
QWEN models are quite nice for local use. Gemini 3 Pro is much better than Codex IMO.
Local? No, not currently. You need about 1TB VRAM. There are many harnesses in development at the time, keep a good look out. Just try many of them, look at the system prompts in particular. Consider DeepSeek using the official API. Consider also tweaking system prompts for whatever tool you end up using. And agree that TUI is meh; we need GUI.
Zed with CC using ACP?
Opencode with CC underneath using Gigacode?
OpenAI codex is also another viable path for what its worth.
I think the best model to my liking open source is kimi k2.5, so maybe you can run that?
Qwen is releasing some new models so I assume keep an eye on those and maybe some model can fit your use case as well?
Zed's ACP client is a wrapper around Agents SDK. So that will be a TOS violation.
You can use Claude CLI as a relay - yes, it needs to be there -but its not that different than use the API
This confirms they're selling those subscriptions at a loss which is simply not sustainable.
They probably are but I don’t think that’s what this confirms. Most consumer flat rate priced services restrict usage outside of the first party apps, because 3rd party and scripted users can generate orders of magnitude more usage than a single user using the app can.
So it makes sense to offer simple flat pricing for first party apps, and usage priced apis for other usage. It’s like the difference between Google Drive and S3.
I get your point - they might count on the user not using their full quota they're officially allowed to use (and if that's the case, Anthropic is not losing money). But then still - IF the user used the whole quota, Anthropic loses.. so what's advertised is not actually honest.
For me, flat rates are simply unfair either ways - if I'm not using the product much, I'm overpaying (and they're ok with that), otherwise it magically turns out that it's no longer ok when I actually want to utilize what I paid for :)
At any rate, they offer the option to be billed exactly on your usage via the API. But if you tell the average person the service costs $x/million tokens they will have no idea how much that actually costs, they won't know what a token is or how much their employees are likely to use. While $30/user/month is something they can easily budget for and get approved.
That's too bad, in a way it was a bit of an unofficial app store for Anthropic - I am sure they've probably looked at that and hopefully this means there's something on it's way.
Not really sure if its even feasible to enforce it unless the idea is to discourage the big players from doing it.
The number one thing we need is cheap abundant decentralized clean energy, and these things are laughable.
Unfortunately neither political party can get all of the above.
Are you implying that no one would use LLM SaaSes and everyone would self-host if energy costs were negligible?
That is...not how it works. People self-hosting don't look at their electricity bill.
I was stuck on the part where they said neither party could provide cheap abundant decentralized clean energy. Biden / Obama did a great job of providing those things, to the point where dirty coal and natural gas are both more expensive than solar or wind.
So, which two parties could they be referring to? The Republicans and the Freedom Caucus?
Bro… re read your comment.
And I just bought my mac mini this morning... Sorry everyone
You know that if you are just using a cloud service and not running local models, you could have just bought a raspberry pi.
Yeah. I know it’s dumb but it’s also a very expensive machine to run BlueBubbles, because iMessage requires a real Mac signed into an Apple ID, and I want a persistent macOS automation host with native Messages, AppleScript, and direct access to my local dev environment, not just a headless Linux box calling APIs.
My M2 macbook pro runs qwen fine, and so will any mac mini with maxed out ram.
The only immediately available was the base 16GB version
Harder to get at the Apple ecosystem. I have an old Macbook that just serves my reminders over the internet.
who knows when Apple decides to enter the game, but they will absolutely crush the personal agent market when they do.
Apple has been doing personal agents for a while. They're crushing it so hard they must be tired of winning at this point.
For instance, the other day, the Siri button in maps told me it couldn't start navigation because it didn't know where it was. It was animating a blue dot with my real time position at the same time.
Don't get me started about the new iOS 26 notification and messaging filters. Those are causing real harm multiple times a day.