Claude Sonnet 3.7

Claude Sonnet 3.7

Quick answer: Claude Sonnet 3.7 (API: claude-sonnet-3-7-20250219) is Anthropic's first model with extended thinking mode, released February 2025. It scores 14.5% on CyberGym, 35.8% on BigCodeBench, and 26.2% on SHADE-Arena — a high SHADE-Arena score indicating robust capability for agentic tasks. Priced at $3 input / $15 output per 1M tokens with a 200K-token context window.

At a Glance

Where Claude Sonnet 3.7 leads

• 26.2% SHADE-Arena overall success — one of the highest scores in its class on this agentic safety evaluation
• 14.5% CyberGym — competitive for a general-purpose model on security tasks
• Extended thinking mode for complex multi-step reasoning
• 200K-token context window
• Multimodal vision support
• Strong SWE-bench Verified performance (62%+ with extended thinking, per Anthropic technical report)

Where it lags

• Superseded by Claude Sonnet 4 for most new deployments
• 35.8% BigCodeBench — moderate coding performance on this benchmark
• Extended thinking adds latency and increases output token cost
• Closed API model with no open weights

Best for: Agentic pipelines requiring reliable extended reasoning; security analysis; coding review workflows; teams already integrated with the Anthropic API.

What Claude Sonnet 3.7 Is

Claude Sonnet 3.7 is the model that introduced extended thinking to the Claude API. Released February 19, 2025, it allows developers to enable visible chain-of-thought reasoning on a per-request basis, with a configurable token budget that balances reasoning depth against latency and cost.

The model's 26.2% SHADE-Arena score is notable: SHADE-Arena measures a model's ability to detect and avoid subtle sabotage scenarios in multi-agent environments, and higher scores indicate stronger agentic situational awareness. This positions Claude Sonnet 3.7 as a particularly capable model for agentic workflows requiring safety and robustness.

Claude Sonnet 3.7's extended thinking capability was its defining feature at launch and drove strong early adoption among teams building coding agents and multi-step reasoning pipelines. The model has since been succeeded by Claude Sonnet 4, which improves on Sonnet 3.7's capabilities.

Specifications

Pricing

Extended thinking tokens are billed as output tokens. Prompt caching provides a 90% discount on cache reads. Pricing per Anthropic pricing page.

Context Window

Claude Sonnet 3.7 has a 200,000-token context window — roughly 150 pages of text in a single request. Extended thinking mode enables up to 64,000 output tokens, allowing for extensive reasoning chains before the final response.

Public Benchmark Scores

Scores from Benchgen evaluations or attributed third-party sources. Well-known launch scores (from Anthropic technical report, February 2025): SWE-bench Verified ~62.3% with extended thinking, GPQA Diamond ~84.8% with extended thinking.

Claude Sonnet 3.7 vs Alternatives

Claude Sonnet 3.7 leads on CyberGym and SHADE-Arena among cost-comparable models, making it the strongest option for agentic safety and security tasks in this price range.

How Claude Sonnet 3.7 Performs on Real Agent Tasks

Claude Sonnet 3.7's 26.2% SHADE-Arena score — the highest among o4-mini, Gemini 2.5 Flash, and comparable peers — indicates strong capability for detecting and avoiding subtle sabotage in multi-agent environments. For agent builders designing pipelines with multiple AI model interactions, this is a meaningful safety signal.

The 14.5% CyberGym score (vs o4-mini's 2.5%) suggests Claude Sonnet 3.7 has meaningfully stronger security analysis capability than OpenAI's cost-tier models. For teams building security-adjacent agentic workflows, this may justify the $3/$15 per 1M price over cheaper alternatives.

Use Claude Sonnet 3.7 via API

import anthropic

client = anthropic.Anthropic(api_key="YOUR_API_KEY")

response = client.messages.create(
    model="claude-sonnet-3-7-20250219",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[{"role": "user", "content": "Analyse this security vulnerability and suggest a fix..."}],
)
print(response.content)

Related Models & Benchmarks

• Compare: Claude 3.7 Sonnet · Claude Sonnet 4 · o4-mini · Gemini 2.5 Pro
• Benchmarks: BigCodeBench · CyberGym · SHADE-Arena

Frequently Asked Questions

Specs and scores sourced from Anthropic's official Claude Sonnet 3.7 announcement (February 2025) and Benchgen evaluations; third-party benchmark scores attributed inline. Pricing cited to the Anthropic pricing page. Last updated 2026-07-23.