Estimate from the version's published specs, benchmarks and reviews, 0-5; scores are only comparable within one criteria year. Scored as launched (Claude 3.5 Sonnet, 20 Jun 2024): 200K-token context and $3 in / $15 out per 1M tokens = $6 blended at 3:1 (both from Anthropic's announcement); text and image input. Anthropic said it runs at twice the speed of Claude 3 Opus but gave no tokens/s, so speed null. Reasoning null: no launch-version Artificial Analysis Intelligence Index entry was opened (AA now shows re-based v4.1 scores); Anthropic's headline GPQA/MMLU/HumanEval figures have no anchor. Coding null (no Terminal-Bench 4.0). Reviews: Willison, Zvi, TechCrunch, The Decoder, Synthedia; these are launch-week analyses with some first-hand use (Willison, Zvi), not systematic tests.[1]
Anthropic released a new model this morning, and I think it's likely now the single best available LLM.
There is a new clear best (non-tiny) LLM. If you want to converse with an LLM, the correct answer is Claude Sonnet 3.5.
But it's more an incremental step than a monumental leap forward.
This makes Sonnet 3.5 potentially the most powerful model available today - at the same price as its predecessor.
Claude 3.5's Sonnet is "more intelligent" than Claude 3's Opus, and it delivers this quality at a far lower price point.
| Claude 3.5 Sonnet · 2024年6月20日 | 0.90 |
|---|---|
| Anthropic claims its latest model is best-in-class · 2024年6月20日 | 0.20 |
| Anthropic launches Claude 3.5, potentially the most capable AI model yet · 2024年6月20日 | 0.80 |
| Anthropic Sonnet 3.5 Sets New Benchmark Standards · 2024年6月21日 | 0.80 |
| On Claude 3.5 Sonnet · 2024年6月24日 | 0.90 |
读者上传的照片,与图钉带来源的媒体分开显示。你可以为吃到或看到的内容评分,在餐厅图钉上还能直接在照片上标记菜品。
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。
Anthropic[2] 写道:“今天,我们发布 Claude 3.5 Sonnet”,并且它“现已在 Claude.ai 上免费提供”;该页面的标注日期现显示为 2024 年 6 月 21 日,但 TechCrunch[4] 和 AWS 均将发布日期报道为 2024 年 6 月 20 日,且模型 ID 为 claude-3-5-sonnet-20240620。
Anthropic's[1] model card addendum: 59.4% on GPQA Diamond, 88.7% on MMLU (5-shot), 92.0% on HumanEval, 71.1% on MATH and 68.3% on MMMU, against Claude 3 Opus's 50.4%, 86.8%, 84.9%, 60.1% and 59.4%; 64% on the internal agentic coding evaluation against Opus's 38%.
"On June 20, 2024, Anthropic[1][2] released Claude 3.5 Sonnet, which, according to the company's own benchmarks, performed better than the larger Claude 3 Opus", alongside Artifacts, which previews code, SVG graphics or websites in a separate window.
TechCrunch, published 2024-06-20T14:00Z: Anthropic[1][2] "is releasing a powerful new generative AI model called Claude 3.5 Sonnet", about twice the speed of Claude 3 Opus, alongside Artifacts. It dates the launch 20 June, where Anthropic's own header now shows the 21 June UTC timestamp.
AWS News Blog, 20 JUN 2024: "Today, Anthropic[1][2] introduced Claude 3.5 Sonnet ... now available in Amazon Bedrock"; 80 percent cheaper than Opus and about 10 percent better than Claude 3 Opus on most vision benchmarks.
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。
原标题: Claude 3.5 Sonnet Released
| 基准测试 | Claude 3.5 Sonnet | Claude 3 Opus | GPT-4o | Gemini 1.5 Pro | Llama-400b(早期快照) |
|---|---|---|---|---|---|
| 研究生水平推理(GPQA, Diamond) | 59.4%*(0-shot CoT) | 50.4%(0-shot CoT) | 53.6%(0-shot CoT) | — | — |
| 本科水平知识(MMLU),5-shot | 88.7%**(5-shot) | 86.8%(5-shot) | — | 85.9%(5-shot) | 86.1%(5-shot) |
| 本科水平知识(MMLU),0-shot CoT | 88.3%(0-shot CoT) | 85.7%(0-shot CoT) | 88.7%(0-shot CoT) | — | — |
| 代码(HumanEval) | 92.0%(0-shot) | 84.9%(0-shot) | 90.2%(0-shot) | 84.1%(0-shot) | 84.1%(0-shot) |
| 多语言数学(MGSM) | 91.6%(0-shot CoT) | 90.7%(0-shot CoT) | 90.5%(0-shot CoT) | 87.5%(8-shot) | — |
| 文本推理(DROP, F1 score) | 87.1(3-shot) | 83.1(3-shot) | 83.4(3-shot) | 74.9(可变样本数) | 83.5(3-shot,预训练模型) |
| 综合评测(BIG-Bench-Hard) | 93.1%(3-shot CoT) | 86.8%(3-shot CoT) | — | 89.2%(3-shot CoT) | 85.3%(3-shot CoT,预训练模型) |
| 数学问题求解(MATH) | 71.1%(0-shot CoT) | 60.1%(0-shot CoT) | 76.6%(0-shot CoT) | 67.7%(4-shot) | 57.8%(4-shot CoT) |
| 小学数学(GSM8K) | 96.4%(0-shot CoT) | 95.0%(0-shot CoT) | — | 90.8%(11-shot) | 94.1%(8-shot CoT) |
* Claude 3.5 Sonnet 在 5-shot CoT、maj@32 条件下的 GPQA 得分为 67.2%。** Claude 3.5 Sonnet 在 5-shot CoT 提示下的 MMLU 得分为 90.4%。
Epoch AI model database (Claude 3.5 Sonnet); Epoch's confidence in the figures: Speculative. Blank tiles are not published.