Skip to main content

​
Overview

Some models generate a separate reasoning stream (sometimes called thinking) in addition to the final answer content. NanoGPT exposes this in an OpenAI-compatible way for Chat Completions:
  • Streaming (SSE): choices[0].delta.reasoning (or legacy choices[0].delta.reasoning_content)
  • Non-streaming: choices[0].message.reasoning (or legacy choices[0].message.reasoning_content)
Not every model exposes reasoning text. If a model does not emit a reasoning stream, these fields may be absent even if the model internally “reasons”. :thinking is model-specific and only works when that exact ID (or a documented alias) exists. -thinking is a legacy alias pattern for some model families only, not universal. Do not assume -thinking works for arbitrary model IDs. Always check GET /api/v1/models for exact valid IDs.

​
Endpoint Variants (Chat Completions)

NanoGPT provides three base paths for chat completions. All accept the same request format and model names, but differ in how reasoning content is delivered: This is the same behavior documented under Reasoning Streams on the Chat Completion page.

​
Controlling Reasoning Output

​
Hide Reasoning

To strip reasoning from the response (both streaming and non-streaming), send:
reasoning.exclude controls output visibility. It is not the same as disabling reasoning compute. Or append the model suffix:
  • :reasoning-exclude
Example:

​
Reasoning Effort

reasoning_effort (or reasoning.effort) controls reasoning depth and also acts as an explicit reasoning-mode signal. An accepted value other than none requests reasoning/thinking behavior. Available levels and defaults depend on the model. Use none to request disabled reasoning only on models that support turning it off. Always-thinking models keep their reasoning requirements; Claude Opus 5.5 rejects attempts to disable thinking.
Or:
Both formats are accepted. If both are present, top-level reasoning_effort is authoritative for Chat Completions request shaping. NanoGPT accepts none, minimal, low, medium, high, xhigh, and max. Not every model supports every level; some routes map an accepted level to the nearest supported effort. max is a separate value from xhigh where the model supports both.

​
Claude Opus 5.5

anthropic/claude-opus-5.5 uses adaptive thinking by default. You do not need a :thinking suffix or an explicit enable flag. Its native effort values are low, medium, high, xhigh, and max; NanoGPT defaults to high when no effort or thinking budget is supplied.
Requests such as reasoning_effort: "none" or thinking.enabled: false return a reasoning_required validation error. To hide returned reasoning while preserving thinking, use reasoning.exclude: true instead. Adaptive thinking also requires compatible tool choice. Use tool_choice: "auto" or "none"; forcing a tool with "required" or a named tool is unsupported for Opus 5.5. In OpenCode, model-level reasoning: true describes a capability to the client. A variant’s reasoningEffort sets the API effort; these settings serve different purposes.

​
Change Effort Within a Conversation

For GPT-6 Astra and Claude Fable 5.1, use inline reasoning effort updates to change effort while keeping the request-level baseline and earlier history unchanged. Updates can precede a user turn or appear at the end of a request to control its generated answer. Chat Completions accepts empty system or developer update messages; Responses and Messages have equivalent forms.

​
Legacy Field Name Compatibility (reasoning_content)

If a client expects reasoning_content instead of reasoning, you can:
  1. Use /api/v1legacy/chat/completions, or
  2. Set reasoning.delta_field: "reasoning_content", or
  3. Use the shorthands reasoning_delta_field / reasoning_content_compat.

​
How Reasoning Appears in Responses

​
Streaming (SSE)

Reasoning deltas appear before or alongside content deltas:

​
Non-Streaming

The final message may include a separate reasoning field:

​
Cost Notes

Reasoning tokens are billed as output tokens. If you enable higher reasoning effort (or use thinking variants), expect higher completion_tokens and higher cost.

​
See Also