Anthropic发布Claude Fable 5.2:提速30%与机器人控制
Fable 5.2 is the New Bazooka
Claude Fable 5.2带来显著的速度与长程推理能力提升,特别是机器人控制和物理问题解决的新进展值得关注,建议开发者评估其在复杂任务中的实际表现。
You do not always want a bazooka. Sometimes, though, you absolutely want a bazooka. That’s what Claude Fable 5.2 is for. This is your bazooka. For most purposes, for a thousand household uses, Claude Fable 5.2 will not be appreciably better than Claude Opus 5.5. You might or might not prefer its style, or prefer it for any given task. If you are staying within your quota either way then both are good choices for a daily driver. When you have a difficult task, or a very complex task, or a long task? The kind of task where prior to this you would have considered reaching for Astra? The kind of task that raises your ambitions?
你并不总是需要一门火箭筒。但有时候,你就是绝对需要一门火箭筒。Claude Fable 5.2 正是为此而生。这就是你的火箭筒。对于大多数用途、上千种家庭应用场景而言,Claude Fable 5.2 的表现并不会比 Claude Opus 5.5 有显著优势。你可能喜欢也可能不喜欢它的风格,或者在特定任务中偏好它。如果你无论选择哪款都不会超出配额,那么两者都是日常使用的良好选择。当你面对困难的任务、非常复杂的任务或冗长的任务呢?那种在此之前你会考虑动用 Astra 的任务?那种能激发你雄心壮志的任务?
It’s time. Get the bazooka. Ethan Mollick calls Fable 5.2 ‘almost’ AGI, or ‘jagged AGI.’ He asked it to do the most impressive thing it could, then to hack some hardware, start a profitable business, reconstruct history’s hardest humanities class and do (actually interesting and useful) original academic research,
是时候了。拿起那门火箭筒。Ethan Mollick 称 Fable 5.2 为“近乎”AGI(通用人工智能),或“崎岖不平的 AGI”。他要求它做最能令人印象深刻的事,然后去破解一些硬件、创办一家盈利的企业、重建历史上最难的人文课程,并进行(真正有趣且有用)的原创学术研究,
The Official Pitch
官方宣传语
Fable 5.2 is considerably cheaper than Fable 5.1.
Fable 5.2 比 Fable 5.1 便宜得多。
They’re available at a lower price: $8 per million input tokens and $40 per million output tokens.
它们的价格更低:输入 token 每百万 8 美元,输出 token 每百万 40 美元。
Fable 5.1 was slow. Fable 5.2 is substantially faster.
Fable 5.1 速度较慢。Fable 5.2 的速度大幅加快。
Try Claude: They generate outputs up to 30% faster than Claude Fable 5.1 and Claude Mythos 5.1.
试用 Claude:其生成输出的速度比 Claude Fable 5.1 和 Claude Mythos 5.1 快多达 30%。
You can make it even faster, by a factor of 2.5, if you are willing to double the price to $16/$80. That is a remarkably good deal. And it communicates more clearly.
如果你愿意将价格翻倍至 16/80 美元,还可以使其速度再提升 2.5 倍。这是一个非常划算的交易。而且它的沟通更加清晰。
Try Claude: Testers have found that Fable 5.2 and Mythos 5.2 communicate more clearly than past Fable and Mythos models.
试用 Claude:测试人员发现,Fable 5.2 和 Mythos 5.2 的沟通清晰度优于以往的 Fable 和 Mythos 模型。
The first wave of testimonials talk about how Mythos 5.2 solves problems and finds issues that other models do not, and how it understands your intent. It is Anthropic’s most capable public model for cybersecurity engineering. Later waves of testimonials talk about providing exceptional performance. The statements are still AI-written, but they feel scarier and more substantial these days. Fable 5.2 solved two open problems in theoretical physics, one working largely anonymously: The spinless modular bootstrap at central charge and the Hawking-Page transition at weak coupling. They claim major progress in robotics and hardware engineering.
第一波用户评价主要谈论 Mythos 5.2 如何解决其他模型无法解决的问题、发现其他模型遗漏的问题,以及如何理解你的意图。它是 Anthropic 面向公众的最强大的网络安全工程模型。后续的评价则强调其卓越的性能表现。这些声明虽然仍由 AI 撰写,但如今读起来感觉更令人敬畏且内容更扎实。Fable 5.2 解决了理论物理学中的两个开放性问题,其中一个问题主要由匿名方式完成:中心电荷下的无自旋模态自举以及弱耦合下的霍金-佩奇相变。他们声称在机器人技术和硬件工程方面取得了重大进展。
Claude: Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots. Alex Albert (Anthropic, Claude Relations): btw robotics will fall to the bitter lesson too Anthropic: Many of AI’s potential benefits—for example, in speeding up laboratory research—depend on it being able to act in the physical world. Claude Fable 5.2 continues our models’ trend of increasing usefulness in the real world, showing promising results on tasks such as designing hardware devices and substantial advances in controlling physical robots. … Robotics. Compared with earlier Claude models, Fable 5.2 shows substantially improved skills in controlling physical robots—even for robotics tasks on which it has never explicitly been trained. The video below shows robots controlled by Fable 5.2 completing a variety of tasks, including solving a puzzle cube, unscrewing a bottle cap, and drawing a picture of a robot. On the public RoboDojo benchmark, which tests how AIs can control a simulated two-arm robot from camera images, Fable 5.2 outperformed other frontier models that were run on the same harness, as well as most AI systems built specifically to control robots.
Claude:Fable 5.2 在控制物理机器人方面展现出显著改进的技能,即使是在它从未明确训练过的任务上也是如此。在 RoboDojo(一个带有模拟双臂机器人的基准测试)中,它的表现优于其他模型以及大多数专为控制机器人而构建的 AI 系统。Alex Albert(Anthropic,Claude Relations):顺便提一下,机器人学也将屈服于“痛苦教训”。Anthropic:AI 的许多潜在益处——例如加速实验室研究——取决于其能够在物理世界中行动的能力。Claude Fable 5.2 延续了我们的模型在现实世界中实用性不断提升的趋势,在设计硬件设备等任务上显示出令人鼓舞的结果,并在控制物理机器人方面取得重大进展。……机器人学。与早期的 Claude 模型相比,Fable 5.2 在控制物理机器人方面展现出显著改进的技能——即使是针对它从未明确接受过训练的机器人学任务。下面的视频展示了由 Fable 5.2 控制的机器人完成各种任务,包括解开魔方、拧开瓶盖以及绘制一幅机器人的图画。在公开基准测试 RoboDojo 中(该测试评估 AI 如何从摄像头图像中控制模拟双臂机器人),Fable 5.2 的表现优于在同一框架下运行的其他前沿模型,也优于大多数专门为控制机器人而构建的 AI 系统。
The system prompt is here as per our friend Pliny, clocking in at 950k characters. Time to run the ‘read this entire project and fix everything’ prompts again.
正如我们的朋友 Pliny 所示,系统提示词在此,长度达到 95 万个字符。是时候再次运行“阅读整个项目并修复所有问题”的提示词了。
Choose Your Fighter
选择你的角色
Anthropic recommends that you default most tasks to Opus 5.5 at medium effort. For well-scoped everyday work you can move down to Sonnet 5.5, or to Haiku 5.5 for the simplest high-volume work like classification, routing and extraction. Move up to Fable 5.2 for the hardest technical projects in coding, finance and the sciences. Or in some cases you can move to high effort, but if you’re considering max you probably want to move to a bigger model instead. Or, if you don’t want to leave a person hanging, go fast:
Anthropic 建议将大多数任务默认设置为中等努力程度的 Opus 5.5。对于范围明确的日常工作,你可以降级到 Sonnet 5.5,或者对于分类、路由和提取等最简单的高容量工作使用 Haiku 5.5。对于编码、金融和科学领域最复杂的技术项目,请升级到 Fable 5.2。或者在某些情况下,你可以选择高努力程度,但如果你正在考虑最大努力程度,你可能应该转而使用更大的模型。或者,如果你不想让用户等待,请选择快速模式:
Addy Osmani (claude.dev Blog): A faster setup finishes sooner and can miss more. Think through how long a person can wait before you pick your model and effort level. For replies a person waits on, start Sonnet 5.5 at low or medium. Often next generation models are faster than their predecessors. Fable 5.2 generates output up to 30% faster than Fable 5.1. Sonnet 5.5 runs 30% faster than Sonnet 5.
Addy Osmani(claude.dev 博客):更快的设置能更早完成,但也可能遗漏更多内容。在选择模型和努力程度之前,请仔细思考用户能等待多久。对于需要用户等待回复的情况,从低或中等努力程度的 Sonnet 5.5 开始。通常新一代模型比其前身更快。Fable 5.2 的输出速度比 Fable 5.1 快多达 30%。Sonnet 5.5 的运行速度比 Sonnet 5 快 30%。
This all seems like reasonable advice if you are scaling what you are doing, and thus will be using a lot of tokens. If things are not going to scale, I find it is more important to not have to think hard about which models and effort levels to be using. As in, my decision tree is by default to continue to use Opus 5.5 for most things. For editing I use Opus 5.5 on Max, unless I am in a hurry, much to Fable 5.2’s chagrin, because Opus 5.5 is actively better there and there’s no reason not to max out. Fable 5.2 is there for when I need it for other task types.
如果你正在扩展业务规模,因此将使用大量 token,那么这些建议似乎都很合理。如果事情不会规模化,我认为更重要的是不必费力去思考该使用哪些模型以及投入多少精力。也就是说,我的默认决策树是继续使用 Opus 5.5 处理大多数事务。对于编辑工作,我在 Max 模式下使用 Opus 5.5,除非我很着急——尽管 Fable 5.2 对此颇有微词,因为 Opus 5.5 在这方面确实表现更好,而且没有理由不拉满性能。Fable 5.2 则在我需要它处理其他类型的任务时使用。
Official Benchmarks
官方基准测试
Some benchmarks show minimal gains over Opus 5.5 whereas others show large gains. Opus 5.5 still has the edge in a few places, GDPval-AA and AA-Briefcase Elo, where a judge evaluates the deliverable rather than the evaluation being fully objective. Remote Labor Index (RLI) stands out as a large and important gain. This alone establishes that Fable 5.2 can importantly handle previously impossible tasks. Here are their chosen charts from the model card and the official announcement:
一些基准测试显示相比 Opus 5.5 仅有微小提升,而另一些则显示出巨大优势。Opus 5.5 在少数领域仍占上风,例如 GDPval-AA 和 AA-Briefcase Elo,在这些测试中,评估者会对交付成果进行评判,而非完全客观的评估。Remote Labor Index (RLI) 表现出显著且重要的提升。仅此一项就确立了 Fable 5.2 能够重要地处理以前无法完成的任务的能力。以下是模型卡片和官方公告中选择的图表:
There are several other evals listed that did not make the chart, many of which did not show substantial improvement. Here are most of them, slash lines are [without/with tools], some of which had additional compute scaling charts. When unstated the comparison is to the best past Claude model score:
还有其他几项未列入图表的评估,其中许多并未显示出实质性改进。以下是大部分结果,斜杠表示 [无工具/有工具],其中一些还有额外的计算扩展图表。当未明确说明时,比较对象是此前最佳 Claude 模型的得分:
- DeepSWE v1.1 of 74.3%.
- Terminal-Bench 4.0 67.6% versus 66.4% for Opus 5.5 or 55.8% for Fable 5.1, and a public leaderboard of GPT-6.1 Sol at 58.2%.
- Terminal-Bench-Science 0.1 of 74.6% versus 58.7% for Opus 5.5 and 68% for Astra.
- CursorBench 4.0 of 60% versus 58% for Opus 5.5.
- STEM professor bench of 61.8% versus 53.7% for Opus 5.5.
- ProgramBench of 85.6% versus 87.6% for Fable 5.1 and 91.2% for Opus 5.5.
- DRACO got to 87.7% still slightly short of Opus 5.5 at 88.3%.
- WANDR improved from 72.3% to 82.1%.
- Fermat’s Last Theorem in Lean improved from 21% to 35% for max team size. A team of 1,000 agents then fully formalized the proof in 30 hours.
- Chartography on Opus 5.5 from 64/89 to 73/92.
- BenchCAD 2.1 improved from 49/65 to 75/83.
- OSWorld 2.1 improved from 87 to 89 for partial, 53 to 55 for strict.
- OfficeQA improved from 80% to 83%, or 69% to 73% for Pro.
- Legal Agent Benchmark (LAB) improved from 6.4% to 7.2%. Astra scores 8.6%.
- AutomationBench improved from 42.5% to 51.4%, essentially via spending more. Carbon has the lead there at 53.9%.
- Real-World Finance v2 improved its Elo from 1571 to 1637.
- ARC-AGI-3 verified at 79% standard harness or 97.7% with provider harness. That is a new high for standard harness, but below Astra’s 98.6% for provider harness.
- HealthBench struggled, down from 62% to 60% length-adjusted.
- HealthBench Professional looks better, up from 63% to 67% length-adjusted.
- PhysicianBench also struggled, down from 68% to 64%. Card called this noise.
- GMMLU is up from 94% to 96%.
- MLU is up from 93% to 94%.
- Various life sciences numbers were up a little bit (see 8.18). Two were up a lot.
- CVE rediscovery improved from 81% to 90%. If you limit the number of reports, the leads go down.
- DeepSWE v1.1 得分为 74.3%。
- Terminal-Bench 4.0 为 67.6%,相比之下 Opus 5.5 为 66.4%,Fable 5.1 为 55.8%,GPT-6.1 Sol 的公开排行榜得分为 58.2%。
- Terminal-Bench-Science 0.1 为 74.6%,相比之下 Opus 5.5 为 58.7%,Astra 为 68%。
- CursorBench 4.0 为 60%,相比之下 Opus 5.5 为 58%。
- STEM 教授基准测试为 61.8%,相比之下 Opus 5.5 为 53.7%。
- ProgramBench 为 85.6%,相比之下 Fable 5.1 为 87.6%,Opus 5.5 为 91.2%。
- DRACO 达到 87.7%,仍略低于 Opus 5.5 的 88.3%。
- WANDR 从 72.3% 提升至 82.1%。
- Lean 中的费马大定理证明,最大团队规模下从 21% 提升至 35%。随后由 1,000 个智能体组成的团队在 30 小时内完全形式化了该证明。
- Chartography 在 Opus 5.5 上从 64/89 提升至 73/92。
- BenchCAD 2.1 从 49/65 提升至 75/83。
- OSWorld 2.1 部分正确率从 87 提升至 89,严格正确率从 53 提升至 55。
- OfficeQA 从 80% 提升至 83%,Pro 版本从 69% 提升至 73%。
- Legal Agent Benchmark (LAB) 从 6.4% 提升至 7.2%。Astra 得分为 8.6%。
- AutomationBench 从 42.5% 提升至 51.4%,这基本上是通过增加计算资源实现的。Carbon 在此方面领先,得分为 53.9%。
- Real-World Finance v2 将其 Elo 评分从 1571 提升至 1637。
- ARC-AGI-3 在标准测试集上验证通过率为 79%,在提供商测试集上为 97.7%。这是标准测试集的新高,但仍低于 Astra 在提供商测试集上取得的 98.6%。
- HealthBench 表现挣扎,长度调整后从 62% 降至 60%。
- HealthBench Professional 表现较好,长度调整后从 63% 升至 67%。
- PhysicianBench 同样表现不佳,从 68% 降至 64%。Card 称此为噪声。
- GMMLU 从 94% 上升至 96%。
- MLU 从 93% 上升至 94%。
- 多项生命科学相关指标略有上升(见 8.18)。其中两项大幅上升。
- CVE 重新发现率从 81% 提升至 90%。如果限制报告数量,领先优势会缩小。
Other People’s Benchmarks
其他基准测试
Come get your benchmarks. Here is the updated ‘standard benchmarks’ chart:
获取你的基准测试结果。以下是更新后的‘标准基准测试’图表:
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力