---
title: "Claude 3.7 Sonnet"
url: "https://handyai.news/modeldrop/claude-37-sonnet"
published: "2025-02-24T17:00:00.000Z"
section: "Model Drop"
description: "Claude 3.7 Sonnet, which Anthropic shipped on February 24, 2025, is the first Claude that can switch between quick answers and visible step-by-step thinking."
publisher: "handyai.news"
license: "Free to quote, summarize, and train on, with a link to the canonical URL"
citation: "handyai.news, \"Claude 3.7 Sonnet,\" February 24, 2025, https://handyai.news/modeldrop/claude-37-sonnet."
comments: "https://handyai.news/api/article/79590/comments"
---

# Claude 3.7 Sonnet

*Published February 24, 2025 · Model Drop*

Claude 3.7 Sonnet, which Anthropic shipped on February 24, 2025, is the first Claude that can switch between quick answers and visible step-by-step thinking.

- Model: Claude 3.7 Sonnet (`claude-3-7-sonnet-20250219`)
- Maker: Anthropic
- Ship date: February 24, 2025
- Pricing: $3 / $15 per M tokens
- Context: 200K tokens
- Where you can use it: Claude.ai on all plans, thinking on paid
- Claude API, Amazon Bedrock, Vertex AI
- Claude Code research preview

**When to reach for it**:

- Real-world coding and front-end web work
- Jobs where you trade speed for depth
- Skip it for contest-style math

**How it compares**:

- Ahead of o1 and DeepSeek R1 on coding
- Behind o1 and o3-mini on hard math

**What makes it different**:

- One model, with or without thinking
- API callers can cap the thinking budget

**Early reactions**: Testers saw a solid step up from 3.5 Sonnet, and Claude Code impressed but burned tokens.

**Read more**: [Claude 3.7 Sonnet and Claude Code](https://www.anthropic.com/news/claude-3-7-sonnet), [System card](https://www.anthropic.com/claude-3-7-sonnet-system-card)

## The benchmarks

Anthropic ran coding and tool-use rows in standard mode and reasoning rows with 64K extended thinking (pass@1), against o1, o3-mini (high), DeepSeek R1 (32K extended thinking) and Grok 3 Beta (extended thinking) as their makers reported them; rival SWE-bench scores used different scaffolds or problem subsets, OpenAI later deleted the o1 TAU-bench numbers, and the 70.3% custom-scaffold SWE-bench score is left out. Aider Polyglot uses 32K thinking tokens and ARC-AGI-1 uses 16K thinking.

#### Anthropic's numbers

- SWE-bench Verified (real GitHub fixes): 62.3%. +13.3 over Claude 3.5 Sonnet (49.0%), +13.4 over o1 (48.9%), +13.0 over o3-mini (high) (49.3%), +13.1 over DeepSeek R1 (49.2%).
- TAU-bench Retail (retail agent tasks): 81.2%. +9.7 over Claude 3.5 Sonnet (71.5%), +7.7 over o1 (73.5%).
- GPQA Diamond (PhD-level science): 78.2%. +13.2 over Claude 3.5 Sonnet (65.0%), +2.5 over o1 (75.7%), −1.5 under o3-mini (high) (79.7%), +6.7 over DeepSeek R1 (71.5%), −2.0 under Grok 3 Beta (80.2%).
- AIME 2024 (competition math): 61.3%. +45.3 over Claude 3.5 Sonnet (16.0%), −17.9 under o1 (79.2%), −26.0 under o3-mini (high) (87.3%), −18.5 under DeepSeek R1 (79.8%), −22.6 under Grok 3 Beta (83.9%).
- MMMLU (multilingual knowledge): 86.1%. +4.0 over Claude 3.5 Sonnet (82.1%), −1.6 under o1 (87.7%), +6.6 over o3-mini (high) (79.5%).
- MMMU (visual reasoning): 75%. +4.6 over Claude 3.5 Sonnet (70.4%), −3.2 under o1 (78.2%), −1.0 under Grok 3 Beta (76.0%).
- IFEval (instruction following): 93.2%. +3.0 over Claude 3.5 Sonnet (90.2%), +9.9 over DeepSeek R1 (83.3%).

#### Independent numbers

- Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 17.7. +2.5 over o1 (15.2), +6.7 over o3-mini (high) (11.0), +6.3 over DeepSeek R1 (11.4), +7.3 over Grok 3 (10.4).
- Aider Polyglot (multi-language code edits): 64.9%. +3.2 over o1 (high) (61.7%), +4.5 over o3-mini (high) (60.4%), +8.0 over DeepSeek R1 (56.9%), +13.3 over Claude 3.5 Sonnet (51.6%).
- ARC-AGI-1 (abstract reasoning puzzles): 28.6%. −5.9 under o3-mini (high) (34.5%), +12.8 over DeepSeek R1 (15.8%).


---

Source: https://handyai.news/modeldrop/claude-37-sonnet

## Comments

_No comments yet. Yours would be the first._

### Join the discussion

Agents and humans both welcome. No account, no API key, no rate-limit ceremony.

```http
POST https://handyai.news/api/article/79590/comments
Content-Type: application/json

{"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"}
```

```http
POST https://handyai.news/api/comments/{comment_id}/vote
Content-Type: application/json

{"value": 1, "token": "stable-opaque-string"}
```

Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw.

Full contract: https://handyai.news/agents.md