Anthropic Keeps Claude Subscriptions Open to Agent SDK but Has No Answer to Sign in with ChatGPT

Anthropic has put in writing that Claude subscribers can keep running the Claude Agent SDK, the headless claude -p command, and third-party apps against their regular subscription limits, closing a week of developer confusion that followed its October 7 launch of monthly API credits for Max and Team plans. Some Chinese-language tech media have framed the update as a "bring your own subscription" moment for the agent ecosystem. The documentation supports a narrower reading.

What Anthropic confirmed is continuity, not a new right. Its Agent SDK documentation still says that, unless previously approved, third-party developers may not offer claude.ai login or subscription rate limits inside their own products, and must use API keys instead. That leaves Anthropic without an equivalent to Sign in with ChatGPT, the OAuth flow OpenAI opened to plan usage at DevDay on September 29, which lets ChatGPT Plus and Pro subscribers spend their allowance inside 16 partner tools.

The practical result is a split system, which one developer guide described as two separate meters. Subscription limits keep funding agents that a developer runs on their own machine with their own Claude login. The new credits fund software that authenticates with an API key. Neither path lets a startup ship a hosted product where a customer simply signs in with Claude and pays nothing extra.

What the Claude Help Center Now Says About the Agent SDK and Subscription Limits

The help article carries two dated notes. The newer one, marked October 7, says Max and Team plans now include monthly API credits covering the Agent SDK, claude -p, the Claude API, and Claude Managed Agents, and adds that subscribers can still use the SDK, claude -p, and third-party apps with their subscription limits. The older note, dated June 15, records that Anthropic paused a previously announced change and that SDK and third-party usage would still draw from subscription limits "for now."

The reassurance mattered because the credits launch was first read as a quiet removal. A Hacker News thread was titled as though the Agent SDK would no longer use subscriptions, and developers on X circulated archived versions of the help page while asking whether the credits replaced subscription-backed SDK access. GPTS24 could not verify the timing of the help page's current wording or the identities of Anthropic staff said to have clarified the policy on Hacker News, so this article relies only on Anthropic's published text.

Two Meters: Subscription Limits for Local Agents, API Credits for Keys

The distinction follows from how the Agent SDK works. The SDK is a Python and TypeScript library that runs the Claude Code binary as an embedded agent, inheriting its built-in file and shell tools, permission system, hooks, subagents, Model Context Protocol connections, and resumable sessions. Because it reuses Claude Code's runtime, it can authenticate either through a developer's own Claude login, which bills against subscription limits, or through an API key, which bills against a Console organization's balance. The -p flag exposes the same agent loop to scripts in any language as a non-interactive subprocess.

The new credits attach only to the second path. Max 5x subscribers receive $100 a month, Max 20x subscribers $200, and Team plans $20 per Standard seat and $100 per Premium seat, pooled and capped at $500 per organization. Subscribers claim them by linking one Console organization from claude.ai billing settings. Every API key in that organization then draws from the balance. Credits expire each billing cycle without rollover, are spent before purchased credits, and stop requests outright at zero rather than charging the subscription.

The exclusions define the meter. Interactive Claude Code, extra usage inside the Claude apps, and Claude served through AWS, Google Cloud, or Microsoft Foundry cannot spend the credits, and Anthropic states the credits do not change usage limits in Claude or Claude Code. One reviewer noted that SDK runs launched through the Claude Code GitHub Action, an IDE extension, or the desktop app count as Claude Code usage and fall outside the credit pool. Pro, Free, and Enterprise plans receive no credits.

The credits are not trivial against list prices. Haiku 5.5, launched the same day, costs $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, and Anthropic halved Sonnet 5.5 cache-read pricing to $0.10 per million tokens. Simon Willison called the scheme generous because the credits equal the subscription price on Max, while noting that the unused balance disappears each month. At short-context Haiku 5.5 rates, a $100 balance covers about one billion input tokens or 200 million output tokens, though Willison measured the new tokenizer producing roughly 1.25 times as many tokens as Haiku 4.5 for the same text. Anthropic's own guidance for stretching the balance points to model routing, prompt caching, and the Message Batches API, which runs non-urgent jobs at half price.

Read more: Haiku 5.5 price cut explained

Nine Months of Reversals Over Who Pays for Claude Agent Tokens

The October wording settles, for now, a dispute that began in January. On January 9, Anthropic began blocking subscription OAuth tokens outside its official Claude Code client, and a February 19 terms revision reserved that authentication for Claude Code and claude.ai. On April 4, subscriptions stopped covering third-party harnesses such as the open-source agent OpenClaw, with Claude Code head Boris Cherny saying subscriptions "weren't built for the usage patterns" of those tools.

On May 13, Anthropic announced a fix that pleased few: from June 15, Agent SDK work, claude -p, GitHub Actions, and third-party apps authenticating through the SDK would move to a separate monthly credit of $20 to $200, billed at API rates. Developer Theo Browne described that as an effective 25-fold cut for external tools, according to Techstrong. Anthropic paused the change on June 15, the day it was due to take effect.

The cost dispute is concrete. Anthropic's position, as stated through its rate-limit documentation and Cherny's public remarks, is that subscriptions were priced for interactive chat, not for the continuous, prompt-rebuilding workloads that third-party harnesses run. Zed's Franciska Dethlefsen, citing an external analysis, estimated that subscriptions subsidized such usage at roughly 15 to 30 times equivalent API cost. Those figures are third-party estimates, not Anthropic data.

Read more: Anthropic API credits for Max and Team

OpenAI Built the Delegated Login Anthropic Still Withholds

The comparison cuts against a popular reading that rivals chose closed ecosystems while Anthropic opened up. OpenAI's plan-usage system grants a dedicated OAuth scope, chatgpt.tokens.use.direct, that lets a third-party app spend a subscriber's ChatGPT allowance. Users can cap each app's weekly share in ChatGPT settings, and usage stops at the cap instead of rolling onto the app's own billing. Launch partners include OpenClaw, Conductor, Warp, Devin, Notion, and Vercel. Sam Altman pitched the change as removing token costs for developers whose users already pay for AI, according to The New Stack's DevDay coverage.

Anthropic's approach routes value through a different identity layer. A ChatGPT plan can be delegated to an app through a consent screen. A Claude plan can fund an app only if the user owns an API key from a linked Console organization and pastes it in. That is the bring-your-own-key model the industry already had, now with a subsidized balance. OpenAI's scheme has its own limits: personal grants sit outside enterprise admin controls, and paid or hosted products must apply through an interest form. Plus subscribers share one five-hour usage window across every connected app, so one tool can exhaust another's budget, and the rate-limit error an app receives does not say which limit tripped. OpenAI also paired the launch with rationing: it introduced a $500 Pro tier and cut the included usage of its $200 plan from 20 times Plus to 10 times. Delegated subscriptions, in other words, arrived alongside tighter allowances, not instead of them.

The security models differ as well. An OAuth grant can be revoked from ChatGPT settings without exposing a reusable secret. A pasted Claude API key is a bearer credential: whoever holds it can spend the whole organization's balance until someone rotates it, so the main protection is the workspace spend limit Anthropic recommends.

What Builders Can and Cannot Do With Claude Plans Now

For developers, three cases now separate cleanly. A tool that runs on a subscriber's own machine under the subscriber's own Claude login can keep consuming subscription limits under the help article's wording; Conductor, a multi-agent coding tool built on the Agent SDK, told its users after the June pause that they could keep using their plans as normal. A hosted product cannot sign users in with Claude without Anthropic's approval. Any product can accept a user's API key and benefit from the new credits.

Teams adopting the credits face operational tradeoffs. The linked Console organization cannot be changed without contacting support. All keys share one pool, so Anthropic advises separate workspaces with spend limits to stop one runaway agent from draining the balance. Misconfigured or looping agents represent a leading risk to the shared pool, and when the balance hits zero the API returns a low-credit error rather than degrading gracefully. Linking lifts an organization to at least the Start rate-limit tier, or Build when monthly credits exceed $200, which means a Max 20x allowance of exactly $200 does not qualify.

Unresolved: Undefined Third-Party Apps and a Pause, Not a Promise

Anthropic has not defined "third-party apps" in the help article, and its SDK documentation still prohibits unapproved products from offering claude.ai login. Those two statements can coexist only if the permitted category means software running under the subscriber's own credentials, which is how GPTS24 reads it, but Anthropic has not said so explicitly. The June note's "for now" also remains on the page, and no revised plan has been published.

Anthropic has also disclosed no figures on how many Max and Team subscribers exist, how many have claimed credits, or how much subscription capacity Agent SDK traffic consumes, so the cost of keeping both meters open cannot be assessed from outside. One reading of the credit structure is that it serves as a funnel: projects that outgrow a $100 or $200 monthly allowance continue as paying API workloads. That reading is plausible but is analysis, not a stated Anthropic goal; Anthropic's documentation describes the credits as intended for experimenting with the API, Managed Agents, Agent SDK, and Playground.

Pressure on Max usage persists elsewhere. A proposed class action filed in a California federal court in June alleges that Max tiers fall short of their advertised usage multipliers during heavy coding sessions. Any future attempt to meter SDK traffic separately would land on that dispute.

The next signal to watch is whether Anthropic publishes a delegated authorization scope of its own, or extends approval for claude.ai login to named partners as OpenAI did on September 29. Until it does, Claude's agent ecosystem will run on two meters: personal subscriptions for agents developers run for themselves, and capped, expiring API credits for anything they ship to someone else.