LLM stop and finish reasons by provider

Every response says why the model stopped: finish_reason in Chat Completions, status and incomplete_details.reason in Responses, stop_reason at Anthropic, finishReason at Gemini. When Gemini blocks a prompt before answering, the reason is in promptFeedback.blockReason instead. The table has one row for each value a provider documents, in each of those formats. Each row gives the value exactly as a response carries it, what it means, what to do about it, and whether the answer is whole: complete, cut short, a tool call, filtered, refused, paused or failed.

Where the rows come from

Each meaning is quoted from the provider's own page or schema, and the Source column gives that page and the day it was read.

• Anthropic: its stop-reasons page on platform.claude.com, which also says what to do about each value.
• OpenAI: its published OpenAPI description on GitHub. It says what a value means, not what to do.
• Gemini: the FinishReason and BlockReason enums of Google's published generative_service.proto on GitHub. It says what a value means, and what to do only about a blocked prompt.

The outcome is read from each value's meaning, so a cut-short answer can be found whichever provider sent it. These are the same facts the finish-reason explainer reads. To look up one value and see what Salyro does with it, use the explainer.

The rows marked Salyro come from Salyro's own code, not from a provider. Salyro has one rule about a finish reason: on /v1/responses, only a finish_reason of length can end an answer incomplete. How it applies depends on whether the answer is streamed:

• Streamed: length ends the stream with response.incomplete instead of response.completed, whatever the answer holds.
• Not streamed: the response is incomplete only when the answer has text. Its message item is then incomplete too. An answer cut short with empty text or only tool calls has no message item, so it ends completed.

An answer that is not streamed and has neither content nor tool calls is no answer to Salyro, whatever its finish reason. A refusal carried only in refusal, an empty message and a thinking-only answer are all examples. Salyro drops it and answers 502 provider_error instead, on /v1/responses and /v1/chat/completions alike. A streamed answer with nothing in it still ends with its finish reason.

With every other value, an answer with something in it ends completed, streamed or not. Otherwise the value passes through: /v1/chat/completions carries finish_reason as Salyro's routing layer returns it, except for that 502, and Salyro keeps no mapping from one provider's values to another's. So the Through Salyro column gives no Chat Completions value for an Anthropic or Gemini reason. It says the mapping is unknown.

How often it is checked

Every fact has the day it was read. Each is due to be read again 90 days later, and Salyro's data report lists every fact with its re-check date. A fact past its date is flagged, not removed. Who re-reads the pages, and whether 90 days is the right interval, is still an open decision. A change to the data changes this table the next time the site is built. Nothing in the table is typed by hand.

What it does not know

• xAI's values. Its page could not be read, so its row says unknown rather than listing values from memory. xAI speaks the Chat Completions format.
• What to do, where the provider does not say. OpenAI's schema and Gemini's enum give meanings only, so What to do is unknown on their rows. It is not filled in with advice.
• Values added after the day shown. A value missing from the table may be newer than the reading.
• OpenAI's own guides. They could not be read from the build. Only the published schema was.

Search by value or meaning, filter by format or outcome, sort by any column, and download the rows you see as CSV. Each row has its own link: select the value.