Changelog

Changes to how the Avey API behaves, newest first.

This page records changes to API behaviour: new endpoints, new or removed fields, changed defaults, changed limits, and changed status codes. Anything that could break or unblock working code belongs here. Corrections to the documentation do not.

08/11/2026

  • The maximum output is now 16,384 tokens, up from 8,192. Reasoning counts toward it. A chat completion or completion with no max_tokens can now run to 16,384 tokens, so set max_tokens if your code relied on the old ceiling to bound cost or latency. See Models and limits.

27/10/2026

Breaking change

A pinned model id such as olive@<version> returns 404. Send olive. avey/olive, the form GET /models lists, also works.

  • New endpoints. The Responses API now supports function tools and file_search, stored responses (previous_response_id, retrieve, delete, and list input items), explicit compaction (POST /responses/compact), and input token counting (POST /responses/input_tokens). reasoning.effort and reasoning_effort set how much Olive thinks. Files (/files) and vector stores (/vector_stores) are new. See Files, Vector stores, Conversation state, and Tools. Web search is coming soon: Web search.
  • Limits, as served. The context window is 65,536 tokens and the maximum output is 8,192 tokens. The request body limit is 4 MB and uploaded files can be up to 5 MB.
  • Rate limits are continuously refilling buckets, per kind of call. Model calls: 60 per minute and 10,000 per day, with a burst of 20. Reads and deletes: 10 per second. File uploads, downloads, and vector store attaches: 20 per hour and 200 per day. Vector store creation: 50 per day. Model calls also have per-minute input and output token limits, which return 429 with "code": "rate_limit_exceeded" and x-ratelimit-*-tokens headers. See Rate limits.
  • Reasoning is returned in message.reasoning_content. Earlier pages named the field reasoning, which still carries the same text. Read reasoning_content. See Reasoning.
  • Client timeout. Set it to 700 seconds. See Errors.

14/09/2026

  • Authorization: Bearer authentication. Avey accepts the standard OpenAI header, and it is now the documented way to authenticate. Integrations that send the key in the X-API-Key header keep working unchanged; new ones should use bearer. See Authentication and Frameworks.

28/07/2026

  • Avey Olive is available. One model on an OpenAI-compatible surface: POST /chat/completions, POST /completions, POST /responses, and GET /models. Streaming is available on all three completion endpoints. Embeddings, audio, and image generation are disabled.
  • Olive returns its reasoning on the assistant message of a chat completion, alongside content. It is an addition to the OpenAI schema; SDKs that do not know about it ignore it. See Reasoning.
  • Earlier limits, since superseded. Rate limits applied per account and denials returned 429 with Retry-After. The current values are in the 27/10/2026 entry and on Rate limits.
  • Requests are authenticated with an API key. Keys are created on the Avey API platform and are account-scoped. See Authentication.
  • Usage totals are always attached to the final chunk of a stream, whether or not you set stream_options.include_usage. See Streaming.
Last updated