Skip to main content
Some models think through a problem before giving a final answer. This makes them stronger at math, code, and logic-heavy tasks. Models with the reasoning capability (for example Deepshi 3.0 and most frontier models) can return their thinking alongside the answer. See Text models for which models support reasoning.

​
Reading the reasoning trace

On a reasoning model, the assistant message can carry extra fields next to content:
  • message.reasoning: the reasoning text.
  • message.reasoning_details: a structured array of reasoning segments.
from openai import OpenAI

client = OpenAI(base_url="https://api.deepshi.ai/v1", api_key="sk-bf-...")

resp = client.chat.completions.create(
    model="deepshi-3.0",
    messages=[{"role": "user", "content": "What is 15% of 240?"}],
    max_tokens=2048,
)

msg = resp.choices[0].message
print(getattr(msg, "reasoning", None))   # the thinking trace (may be None)
print(msg.content)                        # the final answer
If you only want the final answer, read message.content and ignore the reasoning fields.

​
Give reasoning models enough tokens

Reasoning and the visible answer share the same generation budget. If you set max_tokens too low, a reasoning model can spend its whole budget thinking and return empty content with finish_reason: "length". That is a normal 200, not an error. 
{
  "choices": [{ "index": 0, "finish_reason": "length", "message": { "role": "assistant", "content": "" } }]
}
To avoid it, give reasoning models a generous max_tokens so there’s room for both the thinking and the answer.

​
Streaming

Reasoning works with streaming too. Set "stream": true and read choices[].delta. See the Streaming guide.

​
Best practices

  • Allow ample max_tokens on reasoning models (room for thinking and the answer).
  • Use a reasoning model for math, code, and multi-step problems; a non-reasoning model is faster and cheaper for simple tasks.
  • Both reasoning and answer tokens count toward usage.completion_tokens and your cost.