This document compares three reasoning models from different providers:

• OpenAI o3-mini
• Anthropic Claude 3.7 Sonnet
• DeepSeek R1

OpenAI GPT

We selected o3-mini-2025-01-31 as it is currently OpenAI's most cost-effective reasoning model. For detailed pricing information, please visit OpenAI Pricing.

The model accepts a reasoning_effort parameter with three possible values:

• low
• medium
• high

Anthropic Claude 3.7 Sonnet

For this comparison, we're using claude-3-7-sonnet-20250219, Anthropic's dedicated reasoning model. While it's the most expensive model in our comparison, it offers unique capabilities that justify the cost.

Token Configuration

By default, the model has a maximum token limit of 1024.

To prevent incomplete responses, we increased this limit to 4096.

Operation Modes

The claude-3-7-sonnet-20250219 model offers two distinct operation modes:

1. Standard Mode
   • Functions similarly to previous Claude models
   • Provides direct responses without exposing internal reasoning
2. Extended Thinking Mode
   • Reveals Claude's reasoning process before delivering the final answer

API Example:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const response = await client.messages.create({
  model: "claude-3-7-sonnet-20250219",
  max_tokens: 20000,
  thinking: {
    type: "enabled",
    budget_tokens: 16000,
  },
  messages: [
    {
      role: "user",
      content: "Explain quantum entanglement",
    },
  ],
});

Token Parameters

When using extended thinking mode, configure these key parameters:

budget_tokens

• Defines the maximum tokens allowed for internal reasoning processes
• Must be smaller than max_tokens
• Larger budgets enable more thorough analysis of complex problems
• The model may not consume the entire budget, especially above 32K tokens

max_tokens

• Defines the maximum total tokens for the entire response
• Includes budget_tokens as a subset
• For Claude 3.7 Sonnet: a validation error occurs if prompt tokens + max_tokens exceeds the context window

The extended thinking mode provides visibility into the reasoning process. For detailed examples, please refer to the attached Google Colab notebook.

DeepSeek R1

We're using deepseek-reasoner (also known as DeepSeek R1), which offers the most competitive pricing in our comparison.

DeepSeek provides full OpenAI API compatibility. Their documentation recommends using the OpenAI SDK, which enables seamless integration with LangChain-OpenAI as well.

Similar to Anthropic Claude, DeepSeek R1 exposes its reasoning process in the response body. Access both the reasoning and final content through the API:

reasoning_content = response.choices[0].message.reasoning_content
content = response.choices[0].message.content

{
    "completion_tokens": 1312,
    "prompt_tokens": 21,
    "total_tokens": 1333,
    "completion_tokens_details": {
        "accepted_prediction_tokens": None,
        "audio_tokens": None,
        "reasoning_tokens": 334,
        "rejected_prediction_tokens": None
    },
    "prompt_tokens_details": {
        "audio_tokens": None,
        "cached_tokens": 0
    },
    "prompt_cache_hit_tokens": 0,
    "prompt_cache_miss_tokens": 21
}

Key Findings

• All reasoning and thinking processes are billed as output tokens
• Only Claude and DeepSeek expose their reasoning processes in the response
• OpenAI processes reasoning internally within their system
• In some cases, using HTTP libraries or official SDK is better to extract more metadata

Pricing Comparison (per million tokens)