Reasoning Models

Reasoning Models

Some AI models support extended thinking (chain-of-thought reasoning), where the model shows its reasoning process before providing a final answer.

Supported Models

Model	Provider	Reasoning
o1, o1-mini, o3	OpenAI	Extended thinking
Claude 3.5 Sonnet	Anthropic	Extended thinking
DeepSeek R1	DeepSeek	Chain-of-thought
Gemini 2.0 Flash Thinking	Google	Thinking mode

Usage

Reasoning models work through the same API. The key difference is that they may emit ThinkingDelta events during streaming:

Rust

let mut stream = client.chat()
    .user("Solve this step by step: What is 127 * 43?")
    .stream()
    .execute_stream()
    .await?;

while let Some(event) = stream.next().await {
    match event? {
        StreamingEvent::ThinkingDelta { text, .. } => {
            // Model's reasoning process
            print!("[thinking] {text}");
        }
        StreamingEvent::ContentDelta { text, .. } => {
            // Final answer
            print!("{text}");
        }
        _ => {}
    }
}

Python

async for event in client.chat() \
    .user("Solve this step by step: What is 127 * 43?") \
    .stream():
    if event.is_thinking_delta:
        print(f"[thinking] {event.text}", end="")
    elif event.is_content_delta:
        print(event.as_content_delta.text, end="")

How It Works

1. The provider manifest declares capabilities.reasoning: true
2. The streaming decoder recognizes thinking-specific events
3. The EventMapper emits ThinkingDelta for reasoning content
4. ContentDelta events contain the final answer

The protocol manifest handles the provider-specific format differences:

• OpenAI o1: Uses internal reasoning tokens
• Anthropic Claude: Uses thinking content blocks
• DeepSeek R1: Uses <think> tags in content

Tips

• Reasoning models generally produce better results for complex tasks
• They use more tokens (reasoning tokens are counted)
• Temperature may be restricted (some reasoning models ignore it)
• Not all providers support reasoning — check capabilities.reasoning