Endpoints

Reasoning

Control how much thinking a reasoning model does with the reasoning parameter, and read the reasoning tokens it used in each response.

Try it in the playground

Reasoning models think before they answer: OpenAI GPT-5 and o-series, Claude with extended thinking, Gemini 2.5, DeepSeek R1, Qwen and others. api14 accepts a single reasoning parameter and converts it to each provider's native setting. Reasoning tokens are billed as output tokens.

Request parameters

Parameter	Type	Description
`reasoning.effort`	string	`"minimal"`, `"low"`, `"medium"` or `"high"`. Used by OpenAI and Grok models, and mapped to a token budget for others.
`reasoning.max_tokens`	integer	An exact thinking budget in tokens. Used by Anthropic, Gemini and Qwen. Use either this or `effort`, not both.
`reasoning.exclude`	boolean	The model still reasons, but the reasoning text is left out of the response.
`reasoning.enabled`	boolean	Turn reasoning on with default settings.
`include_reasoning`	boolean	Legacy switch, the same as `reasoning: {}` or `reasoning.exclude: false`.

{
  "model": "anthropic/claude-sonnet-4.5",
  "messages": [{ "role": "user", "content": "How many r's are in 'strawberry'? Think carefully." }],
  "reasoning": { "max_tokens": 2000 },
  "max_tokens": 4000
}

max_tokens must be larger than the reasoning budget, because it covers both the thinking and the answer.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api14.com/api/v1",
  apiKey: process.env.API14_API_KEY,
});

const result = await client.chat.completions.create({
  "model": "openai/gpt-5-mini",
  "messages": [
    {
      "role": "user",
      "content": "Explain vector databases in two sentences."
    }
  ],
  "reasoning": {
    "effort": "high"
  }
});
console.log(result.choices[0].message.content);

Response

The reasoning text is returned in message.reasoning, separate from the answer. Some models also return structured reasoning_details:

{
  "id": "gen-1760000000-think",
  "model": "anthropic/claude-sonnet-4.5",
  "choices": [
    {
      "index": 0,
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": "There are 3 r's in \"strawberry\".",
        "reasoning": "s-t-r-a-w-b-e-r-r-y: r at position 3, 8 and 9..."
      }
    }
  ],
  "usage": {
    "prompt_tokens": 22,
    "completion_tokens": 312,
    "total_tokens": 334,
    "completion_tokens_details": { "reasoning_tokens": 290 },
    "cost": 0.004746
  }
}

When streaming, reasoning arrives in delta.reasoning chunks before the delta.content chunks.

Notes

OpenAI o-series and GPT-5 models don't return their reasoning text, but you still pay for reasoning_tokens.
In multi-turn tool use with Claude, send reasoning_details back unchanged in the assistant message so the model can continue its reasoning.
Some models (e.g. deepseek/deepseek-r1) always reason. effort doesn't apply to them.
Higher effort costs more and takes longer. Watch completion_tokens_details.reasoning_tokens.

Errors

The same as Chat completions. A budget that is too small or too large for a model may return an upstream 400. See Errors.