GPT-6 tool-calling and reasoning API compatibility

Summary

OpenAI's GPT-6 model family splits tool use across two endpoints. Function calling over Chat Completions is either unavailable or restricted to non-reasoning requests, depending on the model; combining tools with reasoning requires the Responses API. The rules differ per model (Astra, 6.1 Sol, 6 Sol, 6 Luna) for which reasoning_effort values are accepted, what an omitted effort defaults to, and whether sampling parameters are legal. This page records those constraints and how a client library (yoagent) and an agent (yoyo) map — or fail to map — onto them.

Key Points

• Chat Completions does not support function calling for GPT-6 Astra and GPT-6.1 Sol. Tool calling on those models requires the Responses API.
• GPT-6 Sol and GPT-6 Luna do allow function calling on Chat Completions, but only when reasoning_effort is none. Reasoning plus tools still means Responses.
• reasoning_effort: none is rejected (HTTP 400) by Astra and 6.1 Sol. 6.1 Sol additionally rejects minimal.
• Omitted effort defaults to medium for 6.1 Sol, 6 Sol, and 6 Luna.
• When effort is anything other than none, remove temperature, top_p, and top_logprobs from the request.
• Net effect: for every GPT-6 id, "Chat Completions + tools + reasoning" is an unsupported combination. Astra and 6.1 Sol cannot use Chat Completions with tools at all, since the only effort value that would permit it is the value they reject.

Concepts

• Responses API — the endpoint that supports reasoning together with tool calls across the GPT-6 family.
• Chat Completions — legacy endpoint; tool support for GPT-6 is absent or gated on reasoning_effort: none.
• reasoning_effort — per-request reasoning budget; accepted values and defaults vary by model.