Skip to main content

​
Changelog

​
Platform v2.0.0

This release is being stood up and prepared for cutover because it requires a maintenance window to move the database. Alex is coordinating the maintenance window, and once we have that we can give a definitive cutover time — we are aiming for tomorrow evening. The actual updated platform is being stood up this evening in the new environment.
Released: TBD (Oct 16, 2026)
Previous release: v1.8.3 (October 8, 2026) This release improves platform reliability with fair rate limiting and admission control, adds machine-readable error codes with retry guidance, and lets you opt out of storing Responses API results.

​
What’s New

  • Error codes and retry guidance — Rate-limit (429) and temporary-availability (503) errors now include a machine-readable error.code and a Retry-After header that says when to try again.
  • Opt out of storing responses — POST /v1/responses now supports "store": false. Nothing is saved, and the returned ID can’t be retrieved or used as previous_response_id. The default is still "store": true.
  • 30-day retention for stored responses — Stored Responses API results are now kept for 30 days, the same as OpenAI, and then deleted automatically.

​
Changes

  • Reliability and performance improvements — General improvements to platform stability and responsiveness.

​
Bug Fixes

  • Fair rate limiting and admission control — The platform now applies fair rate limiting and admission control to all traffic. Each organization has limits on how many requests it can send per minute and how many it can run at the same time. The platform only admits the traffic it can serve reliably, so no single account can use up the capacity other customers depend on. Requests over a limit, or that arrive while capacity is temporarily full, get a 429 or 503 with a Retry-After header. Rejected requests don’t count against your organization’s rate limit, so retrying doesn’t lock you out.
  • Some streamed responses ended without a finish signal — They now end cleanly with finish_reason: "length" on chat completions and response.incomplete on the Responses API.
  • Output length could exceed a model’s published maximum — Output is now capped at the maximum listed for each model.
  • Some errors returned an HTML page instead of JSON — Every public API endpoint now always returns JSON in the right format (OpenAI or Anthropic).
  • Some Responses API errors used a non-standard format — They now use the standard error format, with error.code and Retry-After.

​
Customer impact

  • We recommend retrying 429 and 503 responses after the time given in Retry-After. Official OpenAI and Anthropic SDKs do this automatically.
  • Once this release is live, stored Responses API results older than 30 days will be deleted, and requests for them will return 404.
  • Otherwise, no action is required. API keys, endpoints, request formats and pricing are unchanged.

​
Known Issues

  • Follow-up questions need the original image in the request — With chat completions and messages, you can ask a text-only follow-up about an earlier image as long as that image is still in the conversation history you send. If the image is removed from the request entirely, the model can no longer see it. Responses conversations chained with previous_response_id keep the image context automatically.