Anthropic
Claude Haiku 5.5
最大上下文
1M
最大输出
128K
发布日期
暂无
claude-fable-5.1Claude Fable 5 是 Anthropic 的 Mythos 级模型,专为高难度推理、软件工程、知识工作、视觉分析和科学研究而设计。它面向长周期智能体任务,能够在复杂工作流中持续规划与执行,分析大型代码库和文档,并通过多步迭代不断优化结果。Claude Fable 5 非常适合需要高级编码、研究、专业分析以及可靠自主任务完成的实际应用。
上下文窗口
1M Token
最大输出
128K Token
发布日期
2026年9月1日
模态
| 输入价格 | 输出价格 | 缓存读取 | 缓存创建 5m |
|---|---|---|---|
| $10/M | $50/M | $0.25/M | $12.5/M |
推理
工具调用
Temperature 参数
附件
知识库
Endpoint 协议
Claude Fable 5.1 结合高难度自适应推理、长程 Agent 编程、多模态文档理解和百万 Token 上下文。
模型始终启用自适应思考,并默认采用高推理强度,适合需要持续分析而非即时回答的问题。
模型可处理整个代码库、编写测试、从失败步骤中恢复,并在长时间会话中持续自主开发。
模型可在百万 Token 上下文中分析文本和图像输入,包括文件或 PDF 中的图示、图表和表格。
Claude Fable 5.1 适合大型软件项目、多阶段知识工作、长时间运行的 Agent,以及技术或商业文档的视觉审阅。
实现覆盖整个代码库的功能、进行详细代码审查、优化性能,并通过模型编写的测试验证结果。
开展深入研究、分析大规模资料,并将发现转换为可供审阅的文档、电子表格或演示文稿。
理解文件或 PDF 中嵌入的图示、图表和表格,并将其中的依据用于法律、金融、分析或架构问题。
curl 'https://us-api.tokenhub.com/v1/messages' \
-X 'POST' \
-H "Authorization: Bearer $TOKENHUB_API_KEY"| 指数分数 | ||
|---|---|---|
| Artificial Analysis 综合能力指数 | Artificial Analysis 综合能力评分 | 58.1 |
| Artificial Analysis 代码能力指数 | Artificial Analysis 软件任务评分 | 75.2 |
指标来源 Artificial Analysis
关于 Claude Fable 5.1 及通过 TokenHub 使用该模型的常见问题。
Claude Fable 5.1 是 Anthropic 当前面向高难度推理和长时程智能体工作的旗舰模型。它接受文本和图像输入,并生成文本输出。
它适合长时间运行的智能体编程、多步骤研究,以及复杂的文档、电子表格和幻灯片工作。
它提供 100 万 token 的上下文窗口、最多 12.8 万个输出 token、视觉能力、工具调用,以及通过 effort 设置控制的始终开启的自适应思考。
Anthropic 将其相对延迟标为较慢,且每 token 价格高于 Claude Opus 5 或 Sonnet 5。自适应思考无法关闭,安全分类器也可能返回拒绝,因此应测试延迟、成本和拒绝处理。
在 TokenHub API 请求中使用模型标识 claude-fable-5-1,并配合 TokenHub 账户中显示的端点和凭据。投入生产前,请测试请求格式和响应处理。
Anthropic 将 Claude Fable 5.1 列为活跃模型,可通过 Claude API、Amazon Bedrock、Google Cloud、Microsoft Foundry 和 Claude Platform on AWS 使用。公布的 Claude API 价格为每百万输入 token 10 美元、每百万输出 token 50 美元、每百万缓存读取 token 0.25 美元;TokenHub 的计费可能不同。
当高难度、长时程任务所需能力超过 Claude Opus 5 在较高 effort 下的评测表现时,可以选择该模型。对于大多数常规工作负载,Anthropic 建议先从 Claude Opus 5 开始;高吞吐任务则可能更适合更快或更便宜的模型。
通过一个 API Key 接入 Claude Fable 5.1 与更多 AI 模型。
Claude Fable 5.1 媒体与演示
经过验证的 Claude Fable 5.1 官方帖子、视频和社区讨论。
X (Twitter)
Claude ·
We’re introducing Claude Fable 5.1 and Claude Mythos 5.1. They're the world’s most advanced models for coding and knowledge work. https://t.co/8P9PSrWPi3
https://pbs.twimg.com/media/HRJlwmVWcAEyYzf.jpg查看原文Claude ·
Fable 5.1 excels at complex, long-running tasks. And its research capabilities offer an early glimpse of how AI models will contribute to scientific progress. https://t.co/Yj0h0C3LdL
https://pbs.twimg.com/media/HRJl0VlW0AAS6DO.jpg查看原文Claude ·
Cache reads with Fable 5.1 cost 75% less than Fable 5’s. This reduces the cost of the model in practice by around 25% for typical workloads, and up to 45% for highly agentic ones. https://t.co/uGjoKHcbOD
https://pbs.twimg.com/media/HRJhw1hXUAA_Uhm.png查看原文Reddit
Introducing Claude Fable 5.1 and Claude Mythos 5.1
ClaudeOfficial ·
查看原文Used Fable 5.1 all day, here's what's actually different and what to try first
FragrantProgress8376 ·
展开全文
everyone's posting benchmarks. dont care. here's what I found after 3 hours with it, some of this will save you time what its good at big messy context. I dumped like 40 files in and asked "where does the money actually get calculated" and it walked back through the call chain instead of guessing from the filename. older models would pick the file called pricing.js and stop saying no. asked it to make a change and it went "you can, but this breaks the thing on line 200 that depends on it". first time a model has pushed back on me and been right long boring docs. contracts, terms of service, that 90 page pdf your accountant sent. ask it "what would you ask about if you were me" instead of "summarise this". way better output it stops. you ask a one line question you get a one line answer. sounds small. it isnt what its still bad at anything where the answer is on a webpage that changed last week. still have to tell it to go search it will still confidently do the wrong thing if your instructions were vague. its not psychic, its just less stupid writing in your voice. its better than before but if you dont give it 3 of your own examples it sounds like a very polite intern stuff worth trying today paste an error you've been ignoring for months. the one you're scared of give it a decision you're stuck on and say "argue the other side properly". its annoyingly good at this tell it "dont explain, just do it" at the top. saves a lot of scrolling the mythos thing is real, same model, fable just has safety stuff on top for dual use. if you're a normal person doing normal work you will not notice the difference what I'd skip dont use it as a search engine and dont ask it to "make this viral". it can't and neither can you anyway I've had 3 coffees and got more done than last week. back to it edit: no im not sponsored. wish I was, lads edit 2: yes its paid tier. do the maths on your hourly rate and stop asking
Fable 5.1 official prompting docs released
BasicsOnly ·
https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1
https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1查看原文YouTube
Introducing Claude Fable 5.1
Anthropic
查看原文Claude Fable 5.1 runs the forecast overnight
Claude
查看原文Debugging across the whole stack with Claude Fable 5.1
Claude
查看原文