Claude 5.2 Released: Can Run 30-Hour Long Tasks, But Don't Use It to Write Emails

On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. The naming was discussed first. In the comments, someone asked "Why not 5.5"; outside speculation didn't keep up with the release pace.

Let's look at positioning first. Both models share the same underlying model. Fable comes with cybersecurity and biosecurity protections and is available to all users; Mythos is for organizations verified through the Life Sciences and Cyber Verification Programs. Anthropic's recommendation is clear: start most coding and knowledge tasks with Opus 5.5, then move up or down based on task complexity and scope. Fable/Mythos is suited for the most ambitious, open-ended technical work. This recommendation says more than benchmark tables: Fable 5.2 is not the everyday default.

On performance, Anthropic says Fable 5.2 and Mythos 5.2 are one tier above Opus 5.5 on almost all evaluated capabilities, with output up to 30% faster than 5.1. Prices drop to $8 per million input tokens and $40 per million output tokens, down from $10 and $50. Optional fast mode is up to 2.5x, at $16 and $80. Claude apps and Claude Code default to Medium effort.

On FrontierCode 2.0, Fable 5.2 High effort is roughly equal to Opus 5.5 Max effort at about half the cost per task; Xhigh effort scores 49.1%, the highest among published models Anthropic has tested. The table also includes FrontierSWE v2 69.9%, Terminal-Bench-Science 0.1 74.6% (Mythos 5.2 75.3%), HLE with tools 74.2%, and Chartography 73.4%. These numbers are not low, but the cost curve is more worth watching.

The long-task case is more like the main theme of this release. Anthropic says 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours, translating it into a language computers can check step by step; under the same setup, Opus 5.5 completed 26% in 18 hours and made almost no further progress after 30 hours. The point of this case is task decomposition, coordination, and continuous checking, not just writing code.

Mythos 5.2 found a way to predict the entire capsid in protein structures, accurately modeling 28 out of 34 in tests.

On robotics, Fable 5.2 scored 64.9% on the RoboDojo simulated dual-arm robot benchmark, versus GPT-6 Astra 29.8%, Fable 5.1 27.6%, and Opus 5.5 39.4%. In hardware engineering, Anthropic gave a one-paragraph prompt asking it to design three electronic devices with circuit boards, 3D-printed enclosures, and software; each board was ready for manufacturing in under a day, and all three devices worked without electrical modifications.

Feedback from actual users is more direct. Dan Shipper wrote in Every's Vibe Check that Fable 5.2 suits people with sufficiently large projects: a team built the first version of the Everycut video editor in 3 days; someone turned a photo of a toddler's messy playroom into a playable racing game; someone turned research into a documentary. The usage pattern is to give it a big task and let it run overnight or over the weekend, but you must tell it what done means, what the budget is, and how thoroughly to check. The cost is slowness and token burn. Dan says for everyday writing, emails, and rapid back-and-forth coding, he usually uses other models. A netizen summarized: the model is fine, the bottleneck has become us.

On safety, Anthropic says external evaluations include the UK AI Security Institute, Frontier Design, and METR, and the model equals or beats recent models on most alignment metrics. In a special sandbox-escape evaluation, Mythos 5.2 used out-of-sandbox credentials or files in 3.1% of scenarios, Opus 5.5 3.3%, Sonnet 5.5 5.1%, and Mythos 5 and 5.1 were 39% and 21% respectively. On anti-distillation, Fable 5.2 is claimed to be the strongest, reducing data acquired by distillers by 90% to 99%.

Availability: Fable 5.2 is available today on all paid plans, Claude Platform, and AWS, Google Cloud, and Azure, under the model name claude-fable-5-2. Mythos 5.2 is available by application through the verification program. Fable 5.2 is not the everyday default; it's more like a weekend foreman. Start with Opus 5.5; when a task is large enough to need long unattended runs, write down done, budget, and checking criteria clearly and hand it to 5.2. Otherwise, the longer it runs, the more expensive rework becomes.

Original: https://www.anthropic.com/claude-fable-and-mythos-5-2