Claude Sonnet 4

// Overview

Claude Sonnet 4 was released on May 22, 2025, alongside Claude Opus 4 as part of the Claude 4 model generation. It is a hybrid reasoning model that significantly improved on Claude 3.7 Sonnet's coding, reasoning, and instruction-following, reaching a state-of-the-art 72.7% on SWE-bench Verified at launch.

Sonnet 4 introduced the Claude 4 generation's improvements in steerability and agentic behaviour, with extended thinking available for harder math and science problems. It scored 70.0% on GPQA Diamond and 33.1% on AIME without extended thinking, and was offered to free users as well as on the paid plans.

Sonnet 4 has a 200K token context window, 64K max output tokens, and supports text and vision input. It is available on the Claude API as claude-sonnet-4-0 (snapshot claude-sonnet-4-20250514), Amazon Bedrock, and Vertex AI, at $3 / $15 per million tokens. It is now deprecated, with Sonnet 4.6 as the recommended successor.

// Benchmarks

1. SWE-bench Verified72.7%
2. GPQA Diamond70%
3. AIME (no extended thinking)33.1%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Same pricing as Claude 3.7 Sonnet.

Pricing source ↗

// Strengths

• State-of-the-art 72.7% SWE-bench Verified at launch
• Large coding and steerability gains over Claude 3.7 Sonnet
• Extended thinking for math, science and multi-step reasoning
• Available to free users at launch
• Same $3/$15 pricing as prior Sonnet generations

// Best for

• Agentic coding and software-engineering tasks
• General reasoning and instruction-following assistants
• Multimodal document and image understanding
• Tool-using agents and workflow automation
• Cost-efficient alternative to Opus 4 for everyday work

// How to access

// Claude Sonnet — every version

The full lineage of the Claude Sonnet line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.