Anthropic Unveils Claude Fable 5.2 and Mythos 5.2, Restricting Advanced Model to Vetted Organizations
Image to help understand the article
Anthropic unveiled two new artificial intelligence models, Claude Fable 5.2 and Claude Mythos 5.2, on Nov. 9, positioning them for demanding technical applications while limiting access to the more sensitive Mythos model.
The company said the models are designed for complex, open-ended technical work involving engineering, mathematics and robotics. Diane Penn, head of product management for Anthropic's research division, described them as tools intended for experienced users and advanced applications.
Different access rules for the two models
Claude Fable 5.2 is available across platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.2, by contrast, is restricted to organizations that have completed Anthropic's Cyber Verification Program or Life Sciences Verification Program.
Anthropic described Mythos 5.2 as a version of the same underlying model as Fable 5.2 without its safety guardrails. The company said Fable 5.2 blocks approximately 55% fewer requests than its predecessor, Fable 5.1.
Existing subscribers also receive additional access benefits. Pro and Team subscribers can reset their Fable usage limits one to three times during the launch week, and Fable can account for up to 50% of their weekly usage allowance.
Pricing, technical specifications and performance
Standard API pricing for Fable 5.2 is $8 per million input tokens and $40 per million output tokens. Fast mode costs $16 for input and $80 for output. Anthropic said the rates are 20% lower than those of the 5.1 generation. Batch API users receive a 50% discount on both input and output charges.
Mythos 5.2 supports a context window of 1 million tokens and a maximum output of 128,000 tokens. Its default effort setting is high, and its stated knowledge cutoff is June 2026. Anthropic said support for the model will not end before Nov. 9, 2027.
Anthropic's published comparison lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with a medium default effort setting. Sonnet 5.5 costs $2 and $10, respectively, while Haiku 5.5 starts at 10 cents for input and 50 cents for output.
In company-reported benchmarks, Fable 5.2 scored 49.1% on FrontierCode 2.0 using the Xhigh effort setting. The comparison included Opus 5.5 at 45.5% and GPT-6 Astra at 42.8%. Fable 5.1 and Gemini 4 Carbon each scored 41.1% using the medium effort setting.
On FrontierSWE v2, Fable 5.2 recorded 69.9%, ahead of GPT-6 Astra at 65.5%, Opus 5.5 at 62.3%, Gemini 4 Carbon at 61.2% and Fable 5.1 at 56.3%. In the tool-assisted Humanity's Last Exam evaluation, Fable 5.2 scored 74.2%, compared with 67.7% for Opus 5.5.
The company also reported a 64.9% score for Fable 5.2 on RoboDojo Sim, compared with 39.4% for Opus 5.5 and 27.6% for Fable 5.1.
For Mythos 5.2, Anthropic reported a 76.3% score on the cybersecurity evaluation CyScenarioBench. Opus 5.5, measured without safety guardrails, scored 67.6%, while Mythos 5.1 scored 61.7%. On CVE Rediscovery, Mythos 5.2 reached 89.6%, up from 78.3% for its predecessor.
In the Automated Behavioral Audit, Mythos 5.2 recorded an Overall Misaligned Score of 2.67, compared with 2.79 for Opus 5.5, 3.10 for Sonnet 5.5 and 3.28 for Mythos 5.1.
Independent evaluations show strengths and weaknesses
Independent assessments also placed Fable 5.2 near the top of several model rankings, although results varied by task.
Vals AI reported that Fable 5.2 achieved 71.31% accuracy on its Vals Index, ranking first among 36 models, with a testing cost of $17.94 per evaluation. Claude Sonnet 5.5 scored 70.85%, while Gemini 4 Carbon recorded 70.45%.
Fable 5.2 also ranked first on Vals AI's BioMysteryBench, with 83.70% among 31 models, and Vibe Code Bench v1.1, with 93.77% among 127 models. Artificial Analysis assigned Claude Fable 5.2 (Max, Default Fallback) an Intelligence Index score of 65, placing it first among 74 models. Its weighted average cost per index task was $26.54.
However, the model did not lead every category. On the Vals Science Index, it scored 63.74% with a margin of error of 2.30 percentage points, ranking second among 24 models behind GPT-6 Astra at 65.09%. It also placed second on SWE-bench Multimodal, scoring 62.29% with a 2.21-point margin of error, behind Opus 5.5 at 70.21%.
Its performance was weaker on Tax Agent Bench, where it scored 32.09% with a 3.37-point margin of error, ranking 27th among 84 models.
Anthropic highlights software, mathematics and robotics achievements
Anthropic outlined several projects it said demonstrate Fable 5.2's capabilities. In one example, the model migrated an important operational service from Python to Rust, increasing its startup speed by approximately 26 times.
The company also said the model built an operating system capable of booting on a laptop, including disk, keyboard and network drivers and its own core components. In theoretical physics, Anthropic reported that it solved two previously unresolved problems involving black holes.
Another demonstration involved 1,000 Fable 5.2 agents working with minimal human intervention to fully formalize Fermat's Last Theorem in 30 hours. The company also said the model designed three electronic devices with circuit boards, 3D-printed enclosures and software. All three reportedly worked without electrical modifications.
In biological modeling, Fable 5.2 accurately modeled 28 of 34 attempted protein-folding cases. Anthropic also reported that robots controlled by the model completed tasks including solving a puzzle cube and unscrewing bottle caps.
Safety questions remain as Anthropic expands its models
The launch marks Anthropic's first frontier-model release since CEO Dario Amodei called for a slowdown in AI development in September. The company said both models underwent rigorous safety and alignment testing.
However, Anthropic acknowledged that a fundamental challenge remains unresolved: developing evaluations capable of reliably identifying every potential failure before an AI model is deployed.
The distinction between broadly available Fable 5.2 and restricted Mythos 5.2 reflects the company's approach to distributing increasingly capable AI systems. While the former is accessible through major cloud platforms used by American businesses and developers, access to Mythos is limited to organizations that have completed designated verification programs.
For U.S. technology companies and enterprise customers, the release brings new capabilities to familiar cloud infrastructure while highlighting the growing importance of access controls and safety evaluations as advanced AI systems move into engineering, cybersecurity and scientific research.
Comments
Post a Comment