gpt-6.1-sol, the thinking trace and the visible answer share the same max_tokens budget. When the budget is small (300, for example), the model may spend all of it on thinking and get cut off before writing the answer, so the API returns empty_response or an HTTP 200 with an empty message.content.
The new page walks through three checks (finish_reason, reasoning_tokens, max_tokens) and three fixes (raise max_tokens, lower reasoning_effort, switch to streaming), and explains how these requests are billed.
← Back to live updates · 📚 Monthly archive