> ## Documentation Index
> Fetch the complete documentation index at: https://docs.anannas.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Parameters

> Complete parameter reference for chat completions

## Overview

This document describes all parameters available in the `/v1/chat/completions` endpoint. Parameters are validated server-side; unsupported parameters for a given provider are ignored.

## Required Parameters

### `model`

**Type:** `string`\
**Required:** Yes\
**Description:** Model identifier in format `provider/model-name`

Examples:

* `openai/gpt-5-mini`
* `anthropic/claude-3-sonnet`
* `openai/gpt-3.5-turbo`

### `messages`

**Type:** `Message[]`\
**Required:** Yes\
**Minimum:** 1 message\
**Description:** Array of message objects with `role` and `content`

Message roles:

* `system`: System instructions (typically first message)
* `user`: User input
* `assistant`: Model responses (for conversation history)
* `tool`: Tool execution results

## Sampling Parameters

### `temperature`

**Type:** `number`\
**Range:** `0.0` - `2.0`\
**Default:** `1.0`\
**Description:** Controls randomness. Lower values make output more deterministic.

* `0.0`: Most deterministic
* `1.0`: Balanced
* `2.0`: Most random

### `top_p`

**Type:** `number`\
**Range:** `0.0` - `1.0`\
**Default:** `1.0`\
**Description:** Nucleus sampling - considers tokens with cumulative probability mass.

### `max_tokens`

**Type:** `integer`\
**Minimum:** `1`\
**Description:** Maximum tokens to generate. Model-specific limits apply.

### `max_completion_tokens`

**Type:** `integer`\
**Description:** Alternative to `max_tokens` (provider-specific).

### `stop`

**Type:** `string | string[]`\
**Description:** Stop sequences that halt generation. Can be a single string or array.

```json theme={null}
{
  "stop": ["\n\n", "Human:"]
}
```

### `seed`

**Type:** `integer`\
**Description:** Random seed for deterministic outputs. Only supported by some models.

## Penalties

### `frequency_penalty`

**Type:** `number`\
**Range:** `-2.0` - `2.0`\
**Description:** Reduces likelihood of repeating tokens. Positive values decrease repetition.

### `presence_penalty`

**Type:** `number`\
**Range:** `-2.0` - `2.0`\
**Description:** Reduces likelihood of discussing new topics. Positive values encourage new topics.

### `repetition_penalty`

**Type:** `number`\
**Range:** `(0, 2]`\
**Description:** Provider-specific repetition control.

## Advanced Sampling

### `top_k`

**Type:** `integer`\
**Minimum:** `0`\
**Description:** Limits sampling to top K tokens by probability.

<Note>
  **Not allowed with reasoning**: When `reasoning` is enabled, `top_k` is not supported and will result in an error.
</Note>

### `top_a`

**Type:** `number`\
**Range:** `[0, 1]`\
**Description:** Provider-specific sampling parameter.

### `min_p`

**Type:** `number`\
**Range:** `[0, 1]`\
**Description:** Minimum probability threshold for token selection.

### `logit_bias`

**Type:** `{ [token_id: number]: number }`\
**Description:** Bias specific tokens by ID. Values typically range from -100 to 100.

```json theme={null}
{
  "logit_bias": {
    "1234": 10,   // Increase likelihood
    "5678": -10   // Decrease likelihood
  }
}
```

## Tool Calling

### `tools`

**Type:** `Tool[]`\
**Description:** Array of function definitions for tool calling.

```json theme={null}
{
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Get weather for a location",
      "parameters": {
        "type": "object",
        "properties": {
          "location": {"type": "string"}
        },
        "required": ["location"]
      }
    }
  }]
}
```

### `tool_choice`

**Type:** `"none" | "auto" | "required" | { type: "function", function: { name: string } }`\
**Default:** `"auto"`\
**Description:** Controls tool usage behavior.

* `"none"`: No tools called
* `"auto"`: Model decides
* `"required"`: Model must call a tool
* Object: Force specific function

<Note>
  **Restrictions with reasoning**: When `reasoning` is enabled, only `"auto"` or `"none"` are allowed. Using `"required"` or forcing a specific tool will result in an error. This restriction enables interleaved thinking, which allows reasoning between tool calls.
</Note>

### `parallel_tool_calls`

**Type:** `boolean`\
**Default:** `true`\
**Description:** Allow multiple tool calls in a single response.

## Structured Outputs

### `response_format`

**Type:** `{ type: "json_object" } | { type: "json_schema", json_schema: object }`\
**Description:** Enforce JSON output format.

**JSON Object:**

```json theme={null}
{
  "response_format": {
    "type": "json_object"
  }
}
```

**JSON Schema:**

```json theme={null}
{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "response",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "answer": {"type": "string"}
        }
      }
    }
  }
}
```

## Streaming

### `stream`

**Type:** `boolean`\
**Default:** `false`\
**Description:** Enable Server-Sent Events streaming. See [Streaming documentation](/API/Streaming).

## Reasoning

### `reasoning`

**Type:** `object`\
**Description:** Configure reasoning for models that support it.

<Card title="Check Reasoning Support" icon="book">
  For models that support reasoning and their configuration options, visit [anannas.ai/models](https://anannas.ai/models).
</Card>

```json theme={null}
{
  "reasoning": {
    "effort": "high",      // "low" | "medium" | "high"
    "max_tokens": 10000,   // Maximum reasoning tokens
    "enabled": true,       // Enable reasoning
    "exclude": false       // Exclude from response
  }
}
```

<Note>
  **Interleaved Thinking with Tools**

  When using reasoning with tools on Claude 4 models (Sonnet 4.5, Opus 4.5, Haiku 4.5), interleaved thinking is automatically enabled. This allows the model to reason between tool calls. With interleaved thinking, `max_tokens` in the reasoning config can exceed the request's `max_tokens` parameter, as it represents the total budget across all thinking blocks within one assistant turn.

  Interleaved thinking requires `tool_choice: "auto"` (or no tool\_choice specified). Using `tool_choice: "required"` or forcing a specific tool will result in an error when reasoning is enabled.
</Note>

### `thinking_config`

**Type:** `object`\
**Description:** External API alias for `reasoning`. Maps to `reasoning` internally.

```json theme={null}
{
  "thinking_config": {
    "include_thoughts": true,
    "thinking_budget": 10000,
    "thinking_level": "high"
  }
}
```

## Multimodal

### `modalities`

**Type:** `string[]`\
**Description:** Requested output modalities: `["text", "audio", "image"]`

### `audio`

**Type:** `object`\
**Description:** Audio output configuration.

```json theme={null}
{
  "audio": {
    "voice": "alloy",
    "format": "mp3"
  }
}
```

## Prompt Caching

### `prompt_cache_key`

**Type:** `string`\
**Description:** Cache key for OpenAI prompt caching. See [Overview](/API/Overview#prompt-caching).

## Routing

### `models`

**Type:** `string[]`\
**Description:** Fallback model list for routing.

### `route`

**Type:** `"fallback"`\
**Description:** Enable smart fallback routing.

### `provider`

**Type:** `object`\
**Description:** Provider selection preferences.

<Card title="Check Provider Pricing" icon="book">
  For current pricing to set `max_price` limits, visit [anannas.ai/models](https://anannas.ai/models).
</Card>

```json theme={null}
{
  "provider": {
    "order": ["openai", "anthropic"],
    "allow_fallbacks": true,
    "require_parameters": false,
    "data_collection": "deny",
    "zdr": true,
    "only": ["openai"],
    "ignore": ["anthropic"],
    "quantizations": ["q4", "q8"],
    "sort": "price",
    "max_price": {
      "prompt": 0.001,
      "completion": 0.002
    }
  }
}
```

### `fallbacks`

**Type:** `Array<string | object>`\
**Description:** Explicit fallback chain.

```json theme={null}
{
  "fallbacks": [
    "anthropic/claude-3-sonnet",
    {
      "model": "openai/gpt-3.5-turbo",
      "provider": {"only": ["openai"]}
    }
  ]
}
```

## User Tracking

### `user`

**Type:** `string`\
**Description:** Stable identifier for end-users (abuse prevention).

## Metadata

### `metadata`

**Type:** `{ [key: string]: string }`\
**Description:** Custom metadata for request tracking.

```json theme={null}
{
  "metadata": {
    "request_id": "req-123",
    "environment": "production"
  }
}
```

## Provider-Specific Parameters

<Card title="Check Parameter Support" icon="book">
  For detailed parameter support by model and provider, visit [anannas.ai/models](https://anannas.ai/models).
</Card>

Some parameters are provider-specific and may be ignored by others:

* **Mistral**: `safe_prompt`
* **Hyperbolic**: `raw_mode`
* **Grok**: `search_parameters`, `deferred`
* **Anthropic**: `cache_control` in message content

## Parameter Validation

* Invalid parameter values return `400 Bad Request`
* Unsupported parameters are silently ignored
* Provider-specific parameters are passed through when supported

## See Also

* [API Overview](/API/Overview) - Request/response schemas
* [Streaming](/API/Streaming) - Streaming implementation
* [Models](/models) - Model capabilities and parameter support