Reasoning models

Reasoning models, such as GLM-5.3, think before they answer. The thinking counts toward the same token budget as the answer (max_tokens or max_completion_tokens).

The empty answerCopy Anchor Link

With a small budget, the model can spend all of it thinking. The answer then comes back empty:

{
  "choices": [{
    "message": { "role": "assistant", "content": null },
    "finish_reason": "length"
  }]
}

finish_reason: "length" means the budget ran out. Two ways to fix it:

• Turn thinking off for short answers: send "chat_template_kwargs": {"enable_thinking": false}.
• Give it room: raise max_tokens, for example to 2,000 or more for questions that need thought.

Reading the thinkingCopy Anchor Link

When a model shows its thinking, it arrives in reasoning_content (on the message, or on each streamed delta), separate from content. Show it, keep it or drop it as you like. It is billed as output tokens either way.