Amp investigates ChatGPT subscription connection errors
Amp says it is working with OpenAI on ChatGPT connection errors and directs affected users to its legacy connection. Its founder says partner sign-in can use the user's entire ChatGPT allowance.

TL;DR
- Amp is investigating ChatGPT connection errors with OpenAI and directs affected users to the legacy connection, according to its fallback notice.
- Partner sign-in can use 100% of a user's ChatGPT allowance; Amp accepted responsibility for the unclear limit message in a clarification.
- Official partner sign-in draws on existing subscription usage and credits, while remote orbs carry separate Amp charges, as described in Amp's launch thread.
Amp used its own coding agent to rewrite “ChatGPT Sharing Limit Reached.” A separate Windows Codex report describes a login that succeeds through the first model response, then fails when its refresh token is rejected.
Legacy ChatGPT connection
Two ChatGPT connections coexist in Amp's Model Routing settings. Its support replies distinguish:
- Sign in with ChatGPT: the newer partner flow, selected through + Add, as described in a setup reply.
- ChatGPT (legacy): the older connection, also available through + Add, which Amp identified in its fallback instructions.
Work with OpenAI was still ongoing in a subsequent reply. The legacy connection remains Amp's stated workaround for the new connection's failures.
100% ChatGPT allowance
Amp says users can authorize the partner connection to consume their entire allowance in its clarification. It also said OpenAI had delivered the promised behavior and took responsibility for the confusing error.
The consequential detail is the shared pool. OpenAI's usage guide defines an app's limit as a percentage of overall weekly plan usage:
- Eligible app requests consume the existing Codex / ChatGPT work allowance.
- App limits cap consumption without reserving any usage for that app.
- Connecting an app adds no new allowance.
- Spending credits after reaching the plan limit requires a separate permission setting.
Browser accounts and security settings
Amp probed several account conditions during troubleshooting:
- Multiple browser accounts: two ChatGPT accounts active through account switching, raised in a troubleshooting reply.
- Advanced Account Security: a condition Amp said it had previously seen associated with trouble in an earlier reply.
- Daybreak Blue access: queried in one reply, including possible access through a ChatGPT team in a follow-up.
- Recent password resets: another question Amp asked during that exchange.
In the Windows Codex issue, the failing OAuth refresh returned refresh_token_invalidated. The reporter said signing out of ChatGPT's browser accounts and clearing the multi-account session restored normal operation.
Amp pointed affected users toward that browser-session workaround while investigating their connection failure.
Amp fees and abuse checks
Amp worked with OpenAI on the official connection, according to its launch thread. The launch announcement separates subscription-funded inference from paid Megawatt/Gigawatt orbs, Amp's remote machines.
- Local runners: Amp's September 13 Free Agent announcement had already removed the monthly-plan requirement for linked ChatGPT subscriptions. Running agents on users' own computers is free when they supply compute and model access.
- Subscription switched off: Amp invited users to try temporarily switching off their subscription in a launch follow-up. Its routing documentation says deactivating a subscription returns those models to Amp-credit billing unless another connection serves them.
- Automated-abuse limits: Amp retains its own safeguards. It said one user should not have hit those limits and was checking the incident while fixing error handling in a support reply.
Preview API constraints
Amp separately acknowledged a known limitation of the new sign-in in a reply and offered to find a solution. OpenAI's preview documentation specifies concrete restrictions for plan-powered Responses API integrations:
- Request shape: HTTP requests require
store: false,stream: true, and aninputarray. Explicit system-role message items are rejected; instructions or developer messages are supported. - Conversation history: HTTP requests cannot use
previous_response_id. Required history travels ininput; Codex app-server's local history andthread/resumestill work. - Tool execution: hosted Code Interpreter, file search, and native computer use are among the unsupported tools. Local shell and MCP tools can run through supported function/custom tool calls.
- Token renewal: the integrating app must refresh OAuth tokens. With the
env_keysetup, app-server restarts with the renewedACCESS_TOKEN, then resumes the saved thread.