Anthropic发布Claude Fable 5.2登顶全球第一
[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins
全新旗舰模型登顶主流榜单,且揭示了“子代理自检”带来的成本结构变化,对Agent开发者的选型与压测有直接参考价值。
On Friday, we asked Jamin Ball what comes next for Anthropic’s IPO. On Monday the roadshow started and Anthropic shipped the long-anticipated next-version of Claude Fable and Mythos. Of course, it is the new top model in the world, taking back Opus 5.5’s #1 spot from Gemini 4 Carbon which had its brief moment in the sun (10 days).
周五,我们询问了 Jamin Ball Anthropic IPO 之后的下一步计划。周一路演启动,Anthropic 发布了备受期待的 Claude Fable 和 Mythos 新版本。当然,它是全球新的顶级模型,从 Gemini 4 Carbon 手中夺回了 #1 的位置,后者曾短暂地风光过(10天)。
View post on X ↗
在 X 上查看帖子 ↗
Fable comes with a 20% per token price cut… but is more token hungry, as both SemiAnalysis GDPx and Artificial Analysis agree:
Fable 的每个 token 价格降低了 20%……但它更消耗 token,正如 SemiAnalysis GDPx 和 Artificial Analysis 所一致指出的那样:
Fable 5.1 cost $10 per million input tokens and $50 per million output tokens. Fable 5.2 costs $8 and $40. Yet on GDPx, a benchmark of professional tasks, SemiAnalysis paid $88.15 per task, up from $67.16. The run logs give a clue. Where Opus 5.5 and Fable 5.1 would try about 20 test cases, 5.2 sometimes wrote hundreds or thousands and spun up subagents to check its work independently. All that checking helped Fable take first place at 16.4%. You pay for the checks either way.
Fable 5.1 的成本为每百万输入 token 10 美元,每百万输出 token 50 美元。Fable 5.2 的成本分别为 8 美元和 40 美元。然而在 GDPx(专业任务的基准测试)上,SemiAnalysis 每任务花费 88.15 美元,高于之前的 67.16 美元。运行日志提供了一个线索。Opus 5.5 和 Fable 5.1 会尝试约 20 个测试用例,而 5.2 有时编写数百甚至数千个,并启动子代理独立检查其工作。所有这些检查帮助 Fable 以 16.4% 的优势获得第一名。无论如何你都要为这些检查付费。
Which effort setting should you put in production? On FrontierCode 2.0, Anthropic reports that Fable at High matched Opus 5.5 at Max for about half the cost per task… but don’t talk about Fable at Max.
你应该在生产环境中使用哪种努力设置?在 FrontierCode 2.0 上,Anthropic 报告称 Fable 在高(High)模式下与 Opus 5.5 在最大(Max)模式下的表现相当,但每任务成本约为前者的一半……但不要谈论 Fable 在最大(Max)模式下的表现。
AI News for 11/07/2026-11/09/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!
2026年11月7日至11月9日的 AI 新闻。我们检查了 12 个子版块、544 条推文以及没有进一步的 Discord 频道。AINews 的网站允许你搜索所有过往期刊。提醒一下,AINews 现在是 Latent Space 的一个板块。你可以选择订阅或退订电子邮件频率!
AI Twitter Recap
AI Twitter 回顾
Top Story: Claude Fable 5.2 / Mythos 5.2 Benchmarks, Economics, Restrictions, and First Impressions
头条故事:Claude Fable 5.2 / Mythos 5.2 基准测试、经济分析、限制及初步印象
What happened
发生了什么
Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2, claiming both are "a step above Opus 5.5 in almost all the capabilities we evaluated." It tops most independent leaderboards. It also arrives with deliberate RSI nerfs, heavier per-task costs and restricted subscription access.
Anthropic 发布了 Claude Fable 5.2 和 Claude Mythos 5.2,声称两者“在我们评估的大多数能力上都略胜于 Opus 5.5”。它在大多数独立排行榜上名列前茅。同时,它也伴随着刻意进行的 RSI 削弱、更高的每任务成本以及受限的订阅访问权限。
- Launch claims. @claudeai announced the pair as built for "the most daunting technical challenges," with gains in coding, reasoning and vision. @marktenenholtz flagged unusual system-card phrasing: "The model that underlies Claude Fable 5.2 and Claude Mythos 5.2." That implies a single shared base model.
- Vendor-reported headline numbers, per @kimmonismus:
- Up to 30% faster output than the 5.1 models.
- On FrontierCode 2.0, Fable 5.2 at High matched Opus 5.5 at Max for about half the cost per task, and at Xhigh it was the top released model.
- 1,000 Fable 5.2 agents formalized the existing proof of Fermat's Last Theorem in 30 hours. Opus 5.5 completed 26% in 18 hours on the same setup.
- Mythos 5.2 accurately modeled 28 of 34 gene-therapy capsids.
- Alignment was at or above recent Claude models, with external pre-release testing.
- Default effort. Medium is the default effort in the Claude apps and Claude Code.
- Pricing. $8/$40 per 1M input/output tokens, 20% below Fable 5.1's $10/$50 (@jaminball, @arena). The rest of the spec, per Artificial Analysis:
- Cache writes cost $10 (5-minute TTL) or $16 (1-hour TTL); cache reads cost $0.20.
- 1M-token context with text and image input.
- Five effort levels: low, medium, high, xhigh, max.
- Long-horizon demos from Anthropic.
- An OS was built from an empty folder by 1,000+ subagents over 8 days. It boots on a laptop and has its own kernel, disk/keyboard/network drivers, persistent storage, windowing, login, an installer and built-in Claude (@kimmonismus).
- Fable designed three working electronic devices from a short prompt (@kimmonismus).
- It proposed untested chemical synthesis routes.
- It improved on listed solutions to 72 open optimization problems.
- @RLanceMartin published the safety-bounded scaling write-up for the 1,000-agent FLT run.
- Subscription access.
- Ultra and Team Premium users get Fable "resets" scaled by tenure: 1 reset under 6 months, 2 for 6–12 months, 3 above 12 months, usable until March 8 (@ClaudeDevs; echoed by @bcherny and @edwinarbus).
- Pro and Team Standard get a one-week trial capped at half the weekly limit, plus one reset.
- Fable can only consume 50% of a weekly limit.
- Data retention. Fable 5.2 requires data retention by default so Anthropic's safety classifiers can run. Zero data retention is available only to eligible enterprise customers (@code).
- Rollout, day one:
- Cursor (@cursor_ai).
- GitHub Copilot / VS Code, where it "solved more tasks than Fable 5.1 while using fewer steps and tool calls" (@code).
- Perplexity Computer (Pro/Max) and Search (Max) (@perplexity_ai). @AravSrinivas made it his "new default."
- Factory (@factory).
- Hermes Agent via Nous Portal and OpenRouter (@Teknium).
- Arena (@arena) and DesignArena (@grx_xce).
- Official model-selection guide. @ClaudeDevs published guidance on Fable vs Opus vs Sonnet vs Haiku. It covers when to raise effort instead of switching models, and when to push subagents to smaller models.
- 发布声明。@claudeai 宣布这对模型是专为“最艰巨的技术挑战”而构建的,在编码、推理和视觉方面有所提升。@marktenenholtz 指出了系统卡片措辞的不寻常之处:“支撑 Claude Fable 5.2 和 Claude Mythos 5.2 的模型。”这暗示存在一个单一的共享基础模型。
- 供应商报告的 headline 数据,据 @kimmonismus:
- 比 5.1 模型快多达 30%。
- 在 FrontierCode 2.0 上,Fable 5.2 在高(High)模式下与 Opus 5.5 在最大(Max)模式下的表现相当,每任务成本约为前者的一半,而在极高(Xhigh)模式下它成为了已发布的顶级模型。
- 1,000 个 Fable 5.2 代理在 30 小时内形式化了费马大定理的现有证明。在同一设置下,Opus 5.5 在 18 小时内完成了 26%。
- Mythos 5.2 准确模拟了 34 种基因治疗衣壳中的 28 种。
- 对齐效果达到或优于近期的 Claude 模型,并经过了外部预发布测试。
- 默认努力程度(Default effort)。Medium 是 Claude 应用和 Claude Code 中的默认努力程度。
- 定价。每百万输入/输出 token 分别为 $8/$40,比 Fable 5.1 的 $10/$50 低 20% (@jaminball, @arena)。其余规格参数见 Artificial Analysis:
- 缓存写入成本为 $10(TTL 为 5 分钟)或 $16(TTL 为 1 小时);缓存读取成本为 $0.20。
- 支持 1M token 上下文,包含文本和图片输入。
- 五种努力程度:low、medium、high、xhigh、max。
- Anthropic 发布的长周期演示视频。
- 一个操作系统由 1,000 多个子智能体在 8 天内从零文件夹构建而成。它能在笔记本电脑上启动,拥有自己的内核、磁盘/键盘/网络驱动程序、持久化存储、窗口系统、登录功能、安装程序以及内置的 Claude (@kimmonismus)。
- Fable 根据简短提示设计了三种可用的电子设备 (@kimmonismus)。
- 它提出了未经测试的化学合成路线。
- 它对 72 个未解决的优化问题列出的解决方案进行了改进。
- @RLanceMartin 发布了关于 1,000 智能体 FLT 运行的安全边界扩展报告。
- 订阅访问权限。
- Ultra 和 Team Premium 用户获得按使用时长计算的 Fable “重置”次数:使用不足 6 个月获 1 次重置,6–12 个月获 2 次,超过 12 个月获 3 次,有效期至 3 月 8 日 (@ClaudeDevs;@bcherny 和 @edwinarbus 也对此进行了确认)。
- Pro 和 Team Standard 用户可获得为期一周的试用,上限为每周限额的一半,外加一次重置机会。
- Fable 仅能消耗每周限额的 50%。
- 数据保留。Fable 5.2 默认要求保留数据,以便 Anthropic 的安全分类器运行。零数据保留仅适用于符合条件的企业客户 (@code)。
- 上线首日部署情况:
- Cursor (@cursor_ai)。
- GitHub Copilot / VS Code,其中“解决了比 Fable 5.1 更多的任务,且使用的步骤和工具调用更少” (@code)。
- Perplexity Computer (Pro/Max) 和 Search (Max) (@perplexity_ai)。@AravSrinivas 将其设为“新的默认选项”。
- Factory (@factory)。
- 通过 Nous Portal 和 OpenRouter 使用的 Hermes Agent (@Teknium)。
- Arena (@arena) 和 DesignArena (@grx_xce)。
- 官方模型选择指南。@ClaudeDevs 发布了关于 Fable、Opus、Sonnet 和 Haiku 的对比指导。内容涵盖了何时应提升努力程度而非切换模型,以及何时应将子代理推向更小的模型。
Independent evals (facts)
独立评估(事实性)
- Artificial Analysis Intelligence Index v5 (@ArtificialAnlys, breakdown): Fable 5.2 is #1 at 65 (max).
- Fable 5.2 (max) — Index score 65
- Gemini 4 Carbon (high) — Index score 60
- Opus 5.5 (max) — Index score 59
- GPT-6 Astra (max) — Index score 57
- Fable 5.1 (max) — Index score 49
- New records:
- HLE-Diamond — Fable 5.2 80% · Previous best 69%
- AA-Omniscience v1.1 — Fable 5.2 67 · Previous best 54
- Terminal-Bench-Science — Fable 5.2 73% · Previous best 65%
- CritPt v2 hard — Fable 5.2 77% · Previous best 72%
- FrontierCode 2.0 — Fable 5.2 49% · Previous best 46% (Opus 5.5)
- Weaker spots:
- AA-Briefcase: 1785 Elo, within the confidence interval of Opus 5.5 (1800) and behind Sonnet 5.5 (1813).
- GDP.pdf: 24% all-pass vs 26% for Opus 5.5.
- Terminal-Bench 4.0: 60%, level with Opus 5.5 and behind Kimi K3.1 Pro (67%), Gemini 4 Carbon (65%) and Sonnet 5.5 (64%).
- AutomationBench-AA: 76%, ahead of Opus 5.5 (70%) but behind Gemini 4 Carbon (78%).
- Efficiency: At high and xhigh it scores 62 using about 98k output tokens per task, vs 173k for Opus 5.5 max. At max, cost per task is about $26.5, roughly 2.6× Opus 5.5 ($10.0) and 2× Fable 5.1 ($12.99).
- Epoch Capabilities Index (@EpochAIResearch): a record 174, vs 169 for GPT-6 Astra. That is above the pre-Astra reasoning-model trend, which Epoch tied to the AI R&D discussion. Scores came from Epoch's own pre-launch benchmarks.
- ARC-AGI (Verified) (@arcprize):
- ARC-AGI-3, standard harness (new high) — Score 81.2% · Cost $21.9K
- ARC-AGI-3, provider adapter harness — Score 97.7% · Cost $18.6K
- ARC-AGI-2 — Score 93.8% · Cost $0.98/task
- ARC-AGI-1 — Score 98.5% · Cost $0.36/task
- Vals Index (@ValsAI): new #1, leading seven benchmarks across coding, finance and science. "High" was the best-performing effort, above xhigh and max.
- CursorBench (@cursor_ai): 60% at max (top model) and 56.9% at default, vs 49.2% for Fable 5.1. Cursor warns that effort levels behave differently from 5.1, so start at the default rather than carrying over old settings (pricing note).
- Arena Code: WebDev (@arena): #1 at 1832 (xHigh), ahead of Opus 5.5 Max (1818), Gemini 4 Carbon (1768) and GPT-6 Astra (1767). @ml_angelopoulos noted a +87 pt gain over 5.1. Peter Gostev's 100+ generation comparison is here, with 17 sample generations here.
- AA-3D (@ArtificialAnlys): #1 at 1271 Elo (95% CI 1246–1297), with no CI overlap with any other model.
- Runners-up are Sonnet 5.5 (1204) and Opus 5.5 (1202), with GPT-6 Astra at 1168 and Gemini 4 Argon at 1147.
- The score is +132 over Fable 5.1.
- Method (details): each model runs as an agent in an offline sandbox with Blender, Rapier physics, Python and a headless browser, with up to 3h and 250 turns. Humans vote blind on pairs, spending at least 30s per scene.
- VIALS-AA, scientific visual reasoning (@ArtificialAnlys):
- Fable 5.2 scores 67%, vs 64% for Gemini 4 Carbon and 63% for Astra.
- Fable averages about 37 tool calls per task, vs about 8 for Astra. It averages 1.60 calls per turn, the highest on the board (per-turn note).
- Artificial Analysis 智能指数 v5 (@ArtificialAnlys,细分数据):Fable 5.2 以 65 分(满分)位居第一。
- Fable 5.2 (max) — 指数得分 65
- Gemini 4 Carbon (high) — 指数得分 60
- Opus 5.5 (max) — 指数得分 59
- GPT-6 Astra (max) — 指数得分 57
- Fable 5.1 (max) — 指数得分 49
- 新纪录:
- HLE-Diamond — Fable 5.2 达到 80% · 此前最佳为 69%
- AA-Omniscience v1.1 — Fable 5.2 得分为 67 · 此前最佳为 54
- Terminal-Bench-Science — Fable 5.2 达到 73% · 此前最佳为 65%
- CritPt v2 hard — Fable 5.2 达到 77% · 此前最佳为 72%
- FrontierCode 2.0 — Fable 5.2 达到 49% · 此前最佳为 46% (Opus 5.5)
- 较弱环节:
- AA-Briefcase:1785 Elo,处于 Opus 5.5 (1800) 的置信区间内,且落后于 Sonnet 5.5 (1813)。
- GDP.pdf:全部通过率为 24%,而 Opus 5.5 为 26%。
- Terminal-Bench 4.0:60%,与 Opus 5.5 持平,但落后于 Kimi K3.1 Pro (67%)、Gemini 4 Carbon (65%) 和 Sonnet 5.5 (64%)。
- AutomationBench-AA:76%,领先于 Opus 5.5 (70%),但落后于 Gemini 4 Carbon (78%)。
- 效率:在高 (high) 和超高 (xhigh) 设置下,每任务使用约 98k 输出 token 时得分为 62,而 Opus 5.5 max 需 173k。在 max 设置下,每任务成本约为 26.5 美元,大约是 Opus 5.5 (10.0 美元) 的 2.6 倍,也是 Fable 5.1 (12.99 美元) 的 2 倍。
- Epoch 能力指数 (@EpochAIResearch):创下 174 的新高,而 GPT-6 Astra 为 169。该分数高于 Astra 发布前的推理模型趋势,Epoch 将其与 AI R&D 讨论相关联。分数来自 Epoch 自身的发布前基准测试。
- ARC-AGI (已验证) (@arcprize):
- ARC-AGI-3,标准框架(新高)— 得分 81.2% · 成本 21,900 美元
- ARC-AGI-3,提供商适配器工具包 — 得分 97.7% · 成本 18,600 美元
- ARC-AGI-2 — 得分 93.8% · 成本 0.98 美元/任务
- ARC-AGI-1 — 得分 98.5% · 成本 0.36 美元/任务
- Vals Index (@ValsAI):新晋 #1,在编程、金融和科学领域的七个基准测试中领先。"High" 是表现最好的配置,优于 xhigh 和 max。
- CursorBench (@cursor_ai):max 模式下得分为 60%,默认模式下为 56.9%,而 Fable 5.1 为 49.2%。Cursor 警告称,effort 级别的行为与 5.1 不同,因此应从默认设置开始,而不是沿用旧设置(定价说明)。
- Arena Code: WebDev (@arena):以 1832 分(xHigh)位居第一,领先于 Opus 5.5 Max(1818)、Gemini 4 Carbon(1768)和 GPT-6 Astra(1767)。@ml_angelopoulos 指出相比 5.1 版本提升了 87 分。Peter Gostev 的 100+ 次生成对比见此处,17 个样本生成见此处。
- AA-3D (@ArtificialAnlys):以 1271 Elo 分(95% CI 1246–1297)位居第一,且置信区间与其他任何模型均无重叠。
- 紧随其后的是 Sonnet 5.5(1204)和 Opus 5.5(1202),GPT-6 Astra 为 1168,Gemini 4 Argon 为 1147。
- 该分数比 Fable 5.1 高出 132 分。
- 方法(详情):每个模型作为智能体在离线沙盒中运行,配备 Blender、Rapier 物理引擎、Python 和无头浏览器,最多运行 3 小时和 250 轮。人类对成对的场景进行盲评,每场景至少花费 30 秒。
- VIALS-AA,科学视觉推理 (@ArtificialAnlys):
- Fable 5.2 得分为 67%,而 Gemini 4 Carbon 为 64%,Astra 为 63%。
- Fable 平均每个任务调用约 37 次工具,而 Astra 约为 8 次。它平均每轮调用 1.60 次,居所有模型之首(每轮调用说明)。
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