A reasoning model gets room to think

Context

A reasoning model counts its thinking against the output limit of the call. The limit in nr-llm was written for answers: 1000 on a new configuration, 100 in the alternative text preset, 60 for an inline suggestion in the cowriter, 10 for a language detection.

Measured on 2026-11-02 with three real calls per case: at a limit of 10, gpt-6-luna returned no text, finish reason length, in two calls and an error in the third. At 100 with an image, gpt-6-luna returned no text in three of three calls and gpt-6-sol in one of three, having spent all 100 on thinking. A text job of nr-repurpose failed the same way at 1000 on gpt-6-sol, needing about 1050.

The failure was far from its cause. The language detection fell back to English without a word. A structured call reported a JSON syntax error on an empty string. Two consumers had already worked around it, each with a number of its own: the cowriter's field overview with 2048 and the repurpose text pipeline with 16000.

Decision

The output limit sent to a reasoning model of OpenAI is at least 16,000. OpenAiProvider sends the larger of the requested limit and OpenAiModelProfiles::REASONING_OUTPUT_FLOOR on every path: completion, tool round on either transport, vision and stream. A call that switches thinking off with an effort of none keeps its limit. So does a model that does not think. It is the provider's correction, because the provider is what knows that this model counts thinking against this parameter; a caller asks for the length of an answer.

The price is that the limit no longer bounds such a model's answer. A configuration with 1000 can produce a longer text on these models, and an inline suggestion is bounded by its prompt and its stop sequence. What is not used is not paid, and a budget in tokens or cost still applies to what is.

An empty answer that hit the limit is an error. EmptyAnswerMiddleware throws OutputBudgetExhaustedException for a response with no text, no tool call and finish reason length, naming the configuration and the limit. This covers what the floor does not: a limit above the floor that thinking still used up, and the other providers, whose thinking was not measured here. It runs inside the cache layers, so nothing empty is stored, and above the fallback chain, because another configuration is no answer to a limit a person has to raise. Telemetry counts the call as a success, so the health score of a provider that answered does not fall. A truncated text is still an answer and passes as before.

The exception is a ProviderException, so a caller that handles a provider failure handles this one. Three places keep what an empty answer did: a language detection falls back to English, a translation revision keeps the first translation, and the closing text of a run at its round limit stays empty, so the run ends at the limit rather than failed. The task module shows the message.

The upper end of a cost estimate uses the same floor, through OutputCeiling: what the provider may spend is what it is quoted.

Consequences

• ●● A language detection, an alternative text and a short suggestion return text on a reasoning model.

• ● A consumer reads one typed reason instead of a JSON syntax error.

• ◑ Max Tokens no longer bounds the answer of a reasoning model below 16,000, and the upper end of a cost estimate rises with it.

• ◑ The default of 1000 on a new configuration stays. It is a sensible limit for a model that does not think, and changing it would change stored expectations for no gain on the others.

• ◑ Vision reports no finish reason, so an empty description on another provider stays empty.

• ◑ Gemini, OpenRouter and Ollama get the typed error and no floor: their behaviour was not measured.

Net Score: +3

Alternatives considered

Adding a fixed headroom to each limit. Rejected: a limit near the model's maximum would be refused by the API, and the provider does not know that maximum.

Raising each small limit at its call site. Rejected: that is what the two consumers did, and the next preset repeats it.

Lowering the effort for a small limit. Rejected: an effort is a choice of the configuration, and a call cannot tell whether the model needs the thinking.