第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
该文章标注日期为“2025 年 9 月 29 日”,并写道“Claude[2] Sonnet 4.5 今天已在各处可用”;文档中的模型页面标注“发布日期:2025 年 9 月 29 日”,TechCrunch[4] 报道该发布“于周一”进行。
Anthropic's[1][3] model page: claude-sonnet-4-5-20250929, 200K context window, 64K max output, $3/$15 per million tokens, extended thinking, text and images in, a January 2025 reliable knowledge cutoff, "Released September 29, 2025".
Anthropic's[1] system card for Claude[2] Sonnet 4.5, "a new hybrid reasoning large language model" (September 2025), with the alignment and safety evaluations behind its ASL-3 release.
TechCrunch, 2025-09-29: "On Monday, Anthropic[1][3] launched a new frontier model called Claude[2] Sonnet 4.5" at Sonnet 4's $3/$15; researcher David Hershey saw it code autonomously for up to 30 hours; it arrives less than two months after Claude Opus 4.1.
CNBC, 2025-09-29: available to all users from the $183 billion startup; Mike Krieger says it will be the default and recommends it for "basically every use case"; it runs autonomously for 30 hours against Opus 4's seven.
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
来自 Anthropic 发布文章;图表数据读取自其基准测试图片,文章未说明上下文窗口或最大输出。 · 来源
Estimate from the version's published specs, benchmarks and reviews, 0-5; scores are only comparable within one criteria year. Scored as launched (Claude Sonnet 4.5, 29 Sep 2025): 200K-token standard window (Anthropic's model page: 200K; a 1M-token configuration existed as a beta, used by Cognition and noted on the launch post, so scored at the standard 200K); price $3 in / $15 out per 1M tokens = $6 blended at 3:1; text and image input. Anthropic: SWE-bench Verified 77.2%, OSWorld 61.4%; its Terminal-Bench figure (Terminus 2 harness) is not Terminal-Bench 4.0, so coding is null. Reasoning null: AA's current page re-scores 4.5 Sonnet on index v4.1 and no launch-version score was opened. Speed null (no tokens/s). Reviews: Willison (preview access), Zvi, CodeRabbit (code-review benchmark), Cognition (vendor, real agent rebuild), Every (writing tests).[1]
My initial impressions were that it felt like a better model for code than GPT-5-Codex, which has been my preferred coding model since it launched a few weeks ago
My tentative evaluation is a qualified yes. This is likely a big leap in some ways. If I had to pick one 'best coding model in the world' right now it would be Sonnet 4.5.
Sonnet 4.5 gets you close to Opus-level performance at a fraction of the price, making it a pragmatic sweet spot for teams reviewing code at scale.
With Sonnet 4.5, we're seeing the biggest leap since Sonnet 3.6 (the model that was used with Devin's GA): planning performance is up 18%, end-to-end eval scores up 12%, and multi-hour sessions are dramatically faster and more reliable.
Then Anthropic dropped Sonnet 4.5, and within 48 hours my workflow collapsed from two chat interfaces into one.
| Claude Sonnet 4.5 is probably the "best coding model in the world" (at least for now) · 2025年9月29日 | 0.80 |
|---|---|
| Rebuilding Devin for Claude Sonnet 4.5: Lessons and Challenges · 2025年9月29日 | 0.80 |
| Claude Sonnet 4.5 Is A Very Good Model · 2025年10月1日 | 0.80 |
| Claude Sonnet 4.5: Better performance but a paradox · 2025年10月3日 | 0.70 |
| We Tested Claude Sonnet 4.5 for Writing and Editing · 2025年10月23日 | 0.80 |
原标题: Claude Sonnet 4.5 Released
| 基准测试 | Claude Sonnet 4.5 | Claude Opus 4.1 | Claude Sonnet 4 | GPT-5 | Gemini 2.5 Pro |
|---|---|---|---|---|---|
| 智能体编程(SWE-bench Verified) | 77.2% / 配合并行测试时计算达 82.0% | 74.5% / 配合并行测试时计算达 79.4% | 72.7% / 配合并行测试时计算达 80.2% | 72.8%(GPT-5)/ 74.5%(GPT-5-Codex) | 67.2% |
| 智能体终端编程(Terminal-Bench) | 50.0% | 46.5% | 36.4% | 43.8% | 25.3% |
| 智能体工具使用(τ2-bench),零售 | 86.2% | 86.8% | 83.8% | 81.1% | — |
| 智能体工具使用(τ2-bench),航空 | 70.0% | 63.0% | 63.0% | 62.6% | — |
| 智能体工具使用(τ2-bench),电信 | 98.0% | 71.5% | 49.6% | 96.7% | — |
| 计算机操作(OSWorld) | 61.4% | 44.4% | 42.2% | — | — |
| 高中数学竞赛(AIME 2025) | 100%(python)/ 87.0%(不使用工具) | 78.0% | 70.5% | 99.6%(python)/ 94.6%(不使用工具) | 88.0% |
| 研究生水平推理(GPQA Diamond) | 83.4% | 81.0% | 76.1% | 85.7% | 86.4% |
| 多语言问答(MMMLU) | 89.1% | 89.5% | 86.5% | 89.4% | — |
| 视觉推理(MMMU validation) | 77.8% | 77.1% | 74.4% | 84.2% | 82.0% |
| 金融分析(Finance Agent) | 55.3% | 50.9% | 44.5% | 46.9% | 29.4% |