Docs/Reasoning

Reasoning

Models that think before they answer

Reasoning models generate internal thinking before the final answer, which substantially improves math, planning, code, and multi-step analysis. Most flagship models in the catalog are reasoning models; a page badge and the catalog flag mark them.

Using a reasoning model

Nothing special is required — call the model normally. Where a model emits visible reasoning traces, they are passed through in the response and billed as output tokens:

python
resp = client.chat.completions.create(
    model="kimi-k3",
    max_tokens=2000,
    messages=[{"role": "user", "content":
        "A train leaves at 14:40 traveling 90 km/h. A second leaves 25 min later at 120 km/h. When does it catch up?"}],
)

Choosing between reasoning and speed

Use case	Pick
Multi-step math, code review, planning	A full reasoning model (e.g. glm-5-3, kimi-k3)
Classification, extraction, chat latency budgets	A flash-tier model (e.g. glm-5-3-flash, deepseek-v4-1-flash)
Hard problems with tight budgets	Reasoning model + a firm max_tokens cap

NoteReasoning tokens count toward max_tokens and are billed at the output rate. If answers come back truncated, raise the cap — the thinking was cut off, not the answer.

Listing reasoning models

bash
curl -s https://tokenhub-staging.coality.io/api/public/models | jq '[.data[] | select(.is_reasoning) | .id]'