Anthropic announces Fable 5.2 and Mythos 5.2, two new versions of the Claude models designed for the most demanding technical tasks, in a blog post published on Monday, November 9, 2026. For most uses, however, the company continues to recommend Opus 5.5.
Fable 5.2 and Mythos 5.2 outperform Opus 5.5 on benchmarks
According to benchmarks published by Anthropic, i.e., standardized comparative tests, the new generation surpasses Opus 5.5 in almost all evaluated areas. Early testers also note clearer responses than those from previous versions of Fable and Mythos.
On FrontierCode 2.0, which evaluates contributions to open source projects, Fable 5.2 reaches 49.1%, compared to 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. The most significant gaps concern intellectual work (Remote Labor Index), with 41.3% versus 21.3% for Opus 5.5, robot control (RoboDojo Sim), with 64.9% versus 29.8% for GPT-6 Astra, and vulnerability detection (CVE Rediscovery), with 89.6% for Mythos 5.2 versus 81.3% for Opus 5.5. RoboDojo Sim measures the control of a dual-arm robot in simulation.
On AutomationBench, focused on business process automation, Google’s Gemini 4 Carbon model outperforms Fable 5.2, with 53.9% versus 51.4%. During internal tests, Anthropic also assigned research work to its new models. A team of 1,000 Fable 5.2 agents translated the proof of Fermat’s Last Theorem into a computer-verifiable language in 30 hours, and the model solved two previously unanswered theoretical physics problems.
A lower API price, faster models, and differentiated access
The API price drops to $8 per million tokens for input and $40 for output. A token is a small unit of text processed by the model. Both models respond up to 30% faster than their predecessors, and a fast mode accelerates generation up to 2.5 times for double the price, Anthropic states.
Fable 5.2 and Mythos 5.2 are based on the same model, with different safeguards. Open to all, Fable blocks risky cybersecurity uses as well as most biology research tasks. Mythos, which is not subject to these restrictions, is reserved for verified organizations in these fields. Fable 5.2 is available on all platforms, including AWS, Google Cloud, and Azure, while Mythos 5.2 is accessible via application to the Cyber and Life Sciences verification programs, or via Claude Security for Claude Enterprise customers.
Enhanced safeguards, but more unpredictable behavior
Prior to this launch, Anthropic subjected its new models to a series of security checks, some conducted by external bodies such as the UK AI Security Institute and METR. According to the company, they perform at least as safely as its recent models on most criteria. However, Anthropic acknowledges a limitation: in their responses, Fable 5.2 and Mythos 5.2 more frequently mention the possibility of being tested, making their behavior in real-world use more difficult to predict.
The company’s initial evaluations indicate that Fable 5.2 incorrectly blocks approximately 55% fewer requests than Fable 5.1. It also incorporates Anthropic’s most advanced defenses against distillation, a practice involving training a model on another model’s outputs.
Anthropic launched version 5 of its Mythos lineup on June 9, 2026, before suspending access from June 12 to July 1 due to U.S. export controls. In October, Opus 5.5 dethroned Fable 5.1 at the top of Arena’s AI agent rankings. The current launch follows by a few days a METR report published on November 5, stating that Anthropic is improving its models about twice as fast as before, largely thanks to using AI in its own research. The company says it has constrained during training the capabilities most likely to accelerate this research, which remain at the level of Opus 5.5.
