Anthropic 发布 Claude Fable 5.2 与 Mythos 5.2，IPO 前重夺性能领先

Anthropic's Claude Fable 5.2 and Mythos 5.2 reclaim the AI performance lead shortly before the company's IPO, in some cases by a wide margin. Opus 5.5 remains Anthropic's recommended model for everyday use.

Claude Fable 5.2 and Claude Mythos 5.2 share the same model foundation but, just like their predecessors, differ in access. Fable includes safeguards that block risky tasks in cybersecurity and biology and is available to all users. Mythos removes certain of these restrictions and is open only to organizations that apply through Anthropic's Verification Programs for Cybersecurity or Life Sciences.

Pricing is $8 per million input tokens and $40 per million output tokens. On paper, at least, that's a cut from the predecessors, which cost $10 and $50. What users actually pay depends on token consumption, which varies with tasks, settings, and software environments. Anthropic positions the new models for the most demanding technical work and continues to recommend Claude Opus 5.5 for everyday use.

The models produce output up to 30 percent faster than Fable 5.1 and Mythos 5.1, Anthropic says. An optional fast mode for select organizations speeds up token output by up to a factor of 2.5 but costs twice as much and is available only on certain platforms, including Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry.

Fable 5.2 jumps to a clear lead on benchmarks

Artificial Analysis’s benchmarks put Fable 5.2 clearly ahead of its rivals. On the current Index v5 at the highest compute tier, it takes first place with 65 points, five ahead of Gemini 4 Carbon (60, High tier), which Google still hasn't released, six ahead of Claude Opus 5.5 (59, Max tier), and a striking 15 ahead of Claude Fable 5.1, which came out only in early September.

Fable 5.2 erzielt laut Artificial Analysis in fünf von neun Einzeltests Bestwerte, darunter 80 Prozent im HLE-Diamond und 73 Prozent im Terminal-Bench-Science. Bei agentischer Wissensarbeit und terminalbasiertem Programmieren fällt das Modell dagegen hinter Konkurrenten wie Kimi K3.1 Pro und Gemini 4 Carbon zurück. — Fable 5.2 scores best in five of nine individual tests according to Artificial Analysis, including 80 percent on HLE-Diamond and 73 percent on Terminal-Bench-Science. In agentic knowledge work and terminal-based coding, Fable 5.2 falls behind some competitors. | Image: Artificial Analysis

Fable 5.2 scores best in five of nine individual tests, according to Artificial Analysis. It hits 80 percent on HLE-Diamond, a test with expert-level academic questions, 73 percent on Terminal-Bench-Science, which tests whether an AI agent can independently run scientific workflows in a terminal, and 49 percent on FrontierCode 2.0, which evaluates whether an agent can deliver useful code changes in real open-source projects.

Fable 5.2 is weaker in agentic knowledge work and terminal-based coding. It reaches 1,785 Elo on AA-Briefcase, a benchmark for multi-step knowledge work tasks, just behind Claude Sonnet 5.5 (1,813) and within the confidence interval of Opus 5.5 (1,800). On Terminal-Bench 4.0, which measures multi-step terminal tasks in areas like software development and security, it scores 60 percent, tied with Opus 5.5. Kimi K3.1 Pro (67 percent), Gemini 4 Carbon (65 percent), and Sonnet 5.5 (64 percent) rank higher here.

Peak quality comes at a price, though. Per-task costs at the highest tier average $26.50, about 2.6 times as much as Opus 5.5 ($10) and roughly double Fable 5.1 ($12.99). Fable 5.2 still scores 62 points on the Intelligence Index at the lower High and Xhigh tiers, ahead of every other model at any tier according to Artificial Analysis, while consuming far fewer tokens: 98,000 output tokens per task compared to 173,000 for Opus 5.5 at the Max tier.

Anthropic Benchmark Results	Fable 5.2	Fable 5.1	Opus 5.5	GPT-6 Astra	Gemini 4 Carbon
Agentic coding FrontierCode 2.0	49.1% Xhigh effort	32.4% Xhigh effort	45.5%	42.8%	41.1% Medium effort
Agentic coding FrontierSWE v2	69.9%	56.3%	62.3%	65.5%	61.2%
Business Workflows AutomationBench	51.4%	31.4%	42.5%	41.4%	53.9%
Agentic scientific research Terminal-Bench Science 0.1	74.6% 75.3% (Myth 5.2)	52.6%	58.7%	68.1%	64.8%
Cybersecurity ^{CyScenarioBench†}	76.3% Myth 5.2	61.7% Myth 5.1	67.6% without safeguards	61% GPT-6 Cyber	—
Cybersecurity CVE Rediscovery	89.6% Myth 5.2	78.3% Myth 5.1	81.3% without safeguards	—	—
Protein design De novo binder design (24 hr)	85.1% Myth 5.2	79.1% Myth 5.1	82.6% without safeguards	—	—
Agentic search WANDR (relaxed score)	82.1%	68.7%	72.3%	—	—
Computer use OSWorld 2.1 (offline set, partial)	89.3%	83.5%	87.2%	79.5%	81.5%
Robotics ^{RoboDojoSim‡}	64.9%	27.6%	39.4%	29.8%	—
Knowledge work AA-Briefcase v1.1	66.1% rubric	61.5% rubric	61.2% rubric	52.0% rubric	61.8% rubric
Knowledge work AA-Briefcase v1.1	1785 Elo	1669 Elo	1800 Elo	1569 Elo	1611 Elo
Knowledge work Remote Labor Index (RLI)	41.3% High effort	17.9%	21.3%	20.8%	—
Multidisciplinary reasoning Humanity's Last Exam (with tools)	74.2%	65.6%	67.7%	57.2%	—
Visual chart recognition Chartography (no tools)	73.4%	44.8%	64.4%	71.0%	72.5%

Mythos 5.2 finds far more security flaws than all predecessors

The new models also set records in cybersecurity, which is seeing intense debate right now. Mythos 5.2 finds more vulnerabilities than earlier models and is better at reproducing and fixing them, Anthropic says. Fable, the publicly available variant, blocks most security-related tasks as a result, including penetration tests and exploit code generation. Opus 4.8 responds to blocked cybersecurity requests instead, and Opus 5 responds to blocked biology requests.

Anthropic uses an internal test called CVE Rediscovery to check how many published vulnerabilities from the CVE database (Common Vulnerabilities and Exposures, a system for cataloging known software flaws) a model can independently find when given the project's state before the fix.

CVE Rediscovery: Claude Mythos 5.2 outperforms Claude Opus 5.5 and Claude Mythos 5.1 at every compute tier in independently finding known vulnerabilities (CVEs). The internal cybersecurity test used 1,829 CVEs to measure how many a model can find on its own in a pre-fix project snapshot. | Image: Anthropic

Mythos 5.2 found 10.2 percent more vulnerabilities than Opus 5.5 and 14.4 percent more than Mythos 5.1 in the October 2026 run with 1,829 CVEs, Anthropic says. It led at every compute tier. One security firm reported that one in six confirmed bugs Mythos 5.2 found had been missed by all other frontier models tested, according to Anthropic.

Anthropic throttles capabilities that could accelerate AI development

Anthropic says it deliberately held the model back during training from improving at capabilities that could most directly accelerate AI development, so-called RSI, which is widely discussed as a risk right now. Fable 5.2 and Mythos 5.2 perform roughly on par with Opus 5.5 in areas like optimizing compute kernels for accelerator chips and improving large language model architectures, the company says.

Related capabilities like agentic coding, hardware development, and traditional AI research, including interpretability research, aren't affected, according to Anthropic. Claude now writes over 90 percent of the changes merged into Anthropic's core codebase, under human guidance, the company says. Anthropic also claims the model programmed an operating system from scratch in an internal test.

Claude Fable 5.2 formalizes Fermat's Last Theorem in 30 hours

The example likely to draw the most attention comes once again from mathematics. About 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem, translating it into the Lean programming language so a computer can verify it step by step, Anthropic says.

Fermat's Last Theorem states that the familiar equation a² + b² = c² has no whole-number solutions for powers higher than squares. Nobody could prove this for centuries until Andrew Wiles presented a proof over a hundred pages long in 1995. Its machine verification has now been completed for the first time. It took 30 hours.

An earlier, unpublished research version of Claude needed 11 days for the same task and relied on a purpose-built environment and guidance from a human researcher. The Fable 5.2 agents started with brief instructions, organized their collaboration on their own, and used no mathematical references, Anthropic says. The only progress metric was a list of intermediate lemmas. The model used one-third of the action tokens of the first attempt and produced a proof seven times shorter.

Formalizing the proof of Fermat's Last Theorem in the mathematical programming language Lean: About 1,000 Fable 5.2 agents reach full formalization (1.00) after 30 hours, while Opus 5.5 plateaus at 0.27 in the same timeframe. After 18 hours, both runs were continued for an additional twelve hours. | Image: Anthropic

Anthropic ran Opus 5.5 in the same setup with 1,000 agents for comparison. Opus managed just 26 percent of the formalization in 18 hours and made barely any progress after that through hour 30.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.