Anthropic发布Claude Fable 5.2与Mythos 5.2
Anthropic Launches Claude Fable 5.2 and Mythos 5.2: Same Model, Different Safeguards
Anthropic推出同模不同护栏的新版本，性能与成本结构有显著变化，做Agent和复杂编码的同学值得重点关注。
Anthropic just released Claude Fable 5.2 and Claude Mythos 5.2. Claude Fable 5.2 is a frontier model from Anthropic for long-horizon coding, research and agentic work. Mythos 5.2 is the same underlying model with fewer cyber and biology safeguards. It beats Opus 5.5 on most evaluations and costs less than the 5.1 models.
Anthropic 刚刚发布了 Claude Fable 5.2 和 Claude Mythos 5.2。Claude Fable 5.2 是 Anthropic 面向长周期编码、研究和智能体工作的前沿模型。Mythos 5.2 是相同的底层模型，但减少了网络安全和生物学的安全限制。它在大多数评估中击败了 Opus 5.5，且成本低于 5.1 系列模型。
TL;DR
TL;DR（太长不看版）
- Size: 1M-token context window, 128K max output.
- Runs on: Hosted API only. Claude Platform, Amazon Web Services, Google Cloud, Microsoft Azure. No open weights.
- Performance: Ahead of Opus 5.5 on almost every benchmark Anthropic reported.
- Best: RoboDojo Sim 64.9%, more than double GPT-6 Astra (29.8%).
- Worst: AutomationBench 51.4%, behind Gemini 4 Carbon (53.9%).
- Bottom line: Best: Opus-level coding at roughly half the cost per task. Worst: Fable still blocks most cybersecurity work.
- 规模：1M-token 上下文窗口，最大输出 128K。
- 运行平台：仅限托管 API。支持 Claude Platform、Amazon Web Services、Google Cloud、Microsoft Azure。无开源权重。
- 性能：在 Anthropic 报告的所有基准测试中，均领先于 Opus 5.5。
- 最佳表现：RoboDojo Sim 64.9%，是 GPT-6 Astra（29.8%）的两倍多。
- 最差表现：AutomationBench 51.4%，落后于 Gemini 4 Carbon（53.9%）。
- 总结：最佳情况：以约一半的任务成本实现 Opus 级别的编码能力。最差情况：Fable 仍会阻止大多数网络安全工作。
What is the difference between Claude Fable 5.2 and Mythos 5.2?
Claude Fable 5.2 和 Mythos 5.2 有什么区别？
They are one model. Fable 5.2 adds classifiers that block high-risk cybersecurity tasks and most research biology. That lets Anthropic offer it to all users.
它们是同一个模型。Fable 5.2 增加了分类器，用于阻止高风险的网络安全任务和大多数研究生物学内容。这使得 Anthropic 能够向所有用户提供服务。
Mythos 5.2 drops those blocks. Access runs through the Cyber Verification Program and the Life Sciences Verification Program.
Mythos 5.2 移除了这些限制。访问权限需通过网络安全验证计划（Cyber Verification Program）和生命科学验证计划（Life Sciences Verification Program）。
When Fable blocks a query, users get an answer from Opus 4.8 (cyber) or Opus 5 (biology). Anthropic reports about 55% fewer blocks than Fable 5.1. That splits into roughly 60% fewer in cybersecurity and 40% fewer in biology.
当 Fable 拦截查询时，用户会从 Opus 4.8（网络安全）或 Opus 5（生物学）获取答案。Anthropic 报告称，拦截次数比 Fable 5.1 减少了约 55%。其中网络安全领域减少约 60%，生物学领域减少约 40%。
How does Claude Fable 5.2 perform on benchmarks?
Claude Fable 5.2 在基准测试中的表现如何？
Anthropic published one table comparing Fable 5.2 with Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon. All scores are self-reported. Key results:
Anthropic 发布了一张表格，将 Fable 5.2 与 Fable 5.1、Opus 5.5、GPT-6 Astra 和 Gemini 4 Carbon 进行了比较。所有分数均为自我报告。关键结果如下：
- FrontierCode 2.0 (agentic coding): 49.1% at Xhigh effort, vs 45.5% for Opus 5.5.
- FrontierSWE v2: 69.9%, vs 65.5% for GPT-6 Astra.
- Remote Labor Index: 41.3%, vs 21.3% for Opus 5.5.
- Humanity’s Last Exam (with tools): 74.2%, vs 67.7% for Opus 5.5.
- OSWorld 2.1 (partial offline set): 89.3%.
- CVE Rediscovery (Mythos 5.2): 89.6%, vs 81.3% for Opus 5.5 without safeguards.
- FrontierCode 2.0（智能体编码）：在高努力程度下得分为 49.1%，而 Opus 5.5 为 45.5%。
- FrontierSWE v2：得分为 69.9%，而 GPT-6 Astra 为 65.5%。
- Remote Labor Index：得分为 41.3%，而 Opus 5.5 为 21.3%。
- Humanity’s Last Exam（使用工具）：得分为 74.2%，而 Opus 5.5 为 67.7%。
- OSWorld 2.1（部分离线数据集）：得分为 89.3%。
- CVE Rediscovery（Mythos 5.2）：得分为 89.6%，而无安全限制的 Opus 5.5 为 81.3%。
Fable 5.2 does not lead everywhere. Gemini 4 Carbon wins AutomationBench (53.9% vs 51.4%). Opus 5.5 holds a higher AA-Briefcase Elo (1800 vs 1785).
Fable 5.2 并非在所有方面都领先。Gemini 4 Carbon 在 AutomationBench 上获胜（53.9% vs 51.4%）。Opus 5.5 在 AA-Briefcase Elo 评分上更高（1800 vs 1785）。
Is Claude Fable 5.2 more efficient than Opus 5.5?
Claude Fable 5.2 比 Opus 5.5 更高效吗？
Effort levels matter here. On FrontierCode 2.0, Fable 5.2 at High effort scored 45.3%. That matches Opus 5.5 at its best (45.5%) for roughly half the cost per task.
努力程度在此处很重要。在 FrontierCode 2.0 上，Fable 5.2 在高努力程度下的得分为 45.3%。这与 Opus 5.5 的最佳表现（45.5%）相当，但任务成本仅为后者的一半左右。
On Humanity’s Last Exam, Fable 5.2 at Low effort matches Opus 5.5’s top score. Anthropic puts that at less than one-tenth of the cost. On Chartography, Medium effort reaches 66.4%, above Opus 5.5 at Max (64.4%).
在“人类终极考试”（Humanity’s Last Exam）中，Fable 5.2 以低投入达到了与 Opus 5.5 相同的最高分。Anthropic 指出其成本不到后者的十分之一。在 Chartography 测试中，中等投入达到了 66.4%，高于 Opus 5.5 在高投入下的得分（64.4%）。
Outputs are up to 30% faster than the 5.1 models. An optional fast mode speeds generation up to 2.5x.
输出速度比 5.1 模型快高达 30%。可选的快速模式可将生成速度提升多达 2.5 倍。
What did Fable 5.2 do in prerelease testing?
Fable 5.2 在预发布测试中表现如何？
Anthropic lists several long-running agent experiments:
Anthropic 列出了几项长期运行的智能体实验：
- Fermat’s Last Theorem: 1,000 Fable 5.2 agents formalized the proof in Lean in 30 hours. Opus 5.5 reached 26% in 18 hours with the same setup.
- Operating system: One run coordinated over 1,000 subagents for 8 days. The result boots on a laptop.
- Hardware: Claude designed 3 working devices with partner Diode. None needed electrical modifications.
- Physics: Fable 5.2 solved 2 open problems on black holes (preprint 1, preprint 2).
- Migration: Agents moved a production service from Python to Rust. It now starts about 26x faster.
- 费马大定理：1,000 个 Fable 5.2 智能体在 30 小时内用 Lean 形式化了证明。在相同设置下，Opus 5.5 在 18 小时内仅达到 26%。
- 操作系统：一次运行协调了超过 1,000 个子智能体，历时 8 天。结果可在笔记本电脑上启动。
- 硬件：Claude 与合作伙伴 Diode 共同设计了 3 款可工作的设备。这些设备均无需进行电气改装。
- 物理学：Fable 5.2 解决了关于黑洞的 2 个开放问题（预印本 1，预印本 2）。
- 迁移：智能体将一项生产服务从 Python 迁移至 Rust。现在其启动速度提高了约 26 倍。
How safe is Claude Fable 5.2?
Claude Fable 5.2 的安全性如何？
Anthropic tested Mythos 5.2, the version without cyber and biology safeguards. On an automated behavioral audit, it scored 2.67 (lower is better). Opus 5.5 scored 2.79.
Anthropic 测试了 Mythos 5.2，即去除了网络安全和生物安全限制的版本。在自动化行为审计中，它得分为 2.67（分数越低越好）。Opus 5.5 得分为 2.79。
In a sandbox-escape evaluation with 1,800+ scenarios, Mythos 5.2 used out-of-sandbox credentials in 3.1% of cases. Opus 5.5 did so in 3.3%; Mythos 5.1 in 21%.
在包含 1,800 多个场景的沙箱逃逸评估中，Mythos 5.2 在 3.1% 的情况下使用了沙箱外凭据。Opus 5.5 的这一比例为 3.3%；Mythos 5.1 为 21%。
Anthropic also states that it limited specific AI R&D skills during training. Kernel optimization and LLM architecture work sit at about Opus 5.5 level. Full details are in the system card.
Anthropic 还指出，它在训练期间限制了特定的 AI 研发技能。内核优化和 LLM 架构工作水平大致与 Opus 5.5 相当。详细信息请参阅系统卡片（system card）。
How much does Claude Fable 5.2 cost and where is it available?
Claude Fable 5.2 的成本是多少？在哪里可用？
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Fast mode costs $16 and $80. The API ID is claude-fable-5-2.
Fable 5.2 每百万输入令牌收费 8 美元，每百万输出令牌收费 40 美元。快速模式分别为 16 美元和 80 美元。API ID 为 claude-fable-5-2。
It is live on the Claude Platform, AWS, Google Cloud and Microsoft Azure. Fast mode covers Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry. Mythos fast mode is limited to Verification Program members.
它已在 Claude Platform、AWS、Google Cloud 和 Microsoft Azure 上线。快速模式覆盖 Claude Code、Claude Platform、Google Cloud 和 Microsoft Foundry。Mythos 快速模式仅限验证计划成员使用。
Anthropic still recommends Opus 5.5 as the daily driver. Fable is positioned for the hardest open-ended technical work.
Anthropic 仍推荐 Opus 5.5 作为日常主力模型。Fable 定位为用于最困难的开放式技术工作。
How does Claude Fable 5.2 compare with competitors?
Claude Fable 5.2 与竞争对手相比如何？
| Feature | Claude Fable 5.2 | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| Developer | Anthropic | Anthropic | OpenAI | |
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Context window | 1M | 1M | 1,050,000 | Not disclosed by Google |
| Max output | 128K | 128K | 128K | Not disclosed |
| Price (in/out per 1M) | $8 / $40 | $4 / $20 | $10 / $50 | $2 / $10 intro, then $4 / $20 |
| Access | General availability | General availability | General availability | Selected cyber defenders only |
| Weights | Closed | Closed | Closed | Closed |
| FrontierCode 2.0* | 49.1% | 45.5% | 42.8% | 41.1% (Medium) |
| FrontierSWE v2* | 69.9% | 62.3% | 65.5% | 61.2% |
| AutomationBench* | 51.4% | 42.5% | 41.4% | 53.9% |
| HLE (with tools)* | 74.2% | 67.7% | 57.2% | Not reported |
| Chartography* | 73.4% | 64.4% | 71.0% | 72.5% |
| 特性 | Claude Fable 5.2 | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| 开发者 | Anthropic | Anthropic | OpenAI | |
| 参数量 | 未披露 | 未披露 | 未披露 | 未披露 |
| 上下文窗口 | 1M | 1M | 1,050,000 | Google 未披露 |
| 最大输出 | 128K | 128K | 128K | 未披露 |
| 价格（每百万输入/输出） | $8 / $40 | $4 / $20 | $10 / $50 | $2 / $10 入门价，随后为 $4 / $20 |
| 访问权限 | 全面可用 | 全面可用 | 全面可用 | 仅限选定的网络防御者 |
| 权重 | 闭源 | 闭源 | 闭源 | 闭源 |
| FrontierCode 2.0* | 49.1% | 45.5% | 42.8% | 41.1% (Medium) |
| FrontierSWE v2* | 69.9% | 62.3% | 65.5% | 61.2% |
| AutomationBench* | 51.4% | 42.5% | 41.4% | 53.9% |
| HLE (带工具)* | 74.2% | 67.7% | 57.2% | 未报告 |
| Chartography* | 73.4% | 64.4% | 71.0% | 72.5% |
*Benchmark rows come from Anthropic’s launch table. Anthropic ran or reported these scores; they are not independently verified.
*基准测试行数据来自 Anthropic 的发布表格。Anthropic 运行或报告了这些分数；它们未经独立验证。
Key Takeaways
关键要点
- Fable 5.2 and Mythos 5.2 are 1 model with 2 safeguard levels.
- 49.1% on FrontierCode 2.0, ahead of Opus 5.5 and GPT-6 Astra.
- $8/$40 per million tokens, with a 2.5x fast mode at $16/$80.
- Fable blocks about 55% fewer requests than Fable 5.1.
- Gemini 4 Carbon still leads AutomationBench; scores are self-reported.
- Fable 5.2 和 Mythos 5.2 是同一个模型，具有两种安全级别。
- 在 FrontierCode 2.0 上得分为 49.1%，领先于 Opus 5.5 和 GPT-6 Astra。
- 每百万令牌价格为 $8/$40，快速模式价格为 $16/$80（2.5倍）。
- Fable 拦截的请求比 Fable 5.1 少约 55%。
- Gemini 4 Carbon 仍在 AutomationBench 中领先；分数为自我报告。
Check out the launch post, the system card and the model docs. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
请查看发布文章、系统卡片和模型文档。此外，欢迎在 Twitter 上关注我们，别忘了加入我们有 15 万+成员的 ML SubReddit 并订阅我们的新闻通讯。等等！你在 Telegram 上吗？现在你也可以在 Telegram 上加入我们。
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