News · 2026-11-10

Anthropic releases Fable and Mythos 5.2 with shared training-time limits

Anthropic has released Fable 5.2 and Mythos 5.2, two access tiers for the same underlying model with different cyber and biology safeguards. The company also says both share a training-time limit on specific AI-development capabilities, making its safety choices part of the capability story.

Key facts

• Anthropic dates the release to November 9, 2026.
• Fable is broadly available; Mythos requires organizational verification.
• ARC Prize reports 97.7% on ARC-AGI-3 using a provider adapter, versus 81.2% with its standard harness.
• Primary source: Anthropic’s announcement.

Anthropic says it limited progress in “specific research capabilities” associated with accelerating AI development, including accelerator-kernel optimization and improvements to language-model architectures. It says these skills remain about at Opus 5.5 level, while adjacent areas including agentic coding and conventional machine-learning research were unaffected. That latter claim is the company's assessment; the source does not publish enough evidence to rule out spillover into nearby tasks.

This is separate from an access guardrail. Think of a training intervention as changing what a worker has learned, while a serving-time guardrail is a gate that decides which requests reach that worker. Anthropic says Fable applies safeguards for high-risk cyber and biology requests, while verified Mythos users have certain safeguards removed. Both models share the training intervention. Anthropic also describes a small set of separate serving-time fallbacks for AI research topics.

The ARC Prize evaluation supplies an independent check with an important qualification. Its 97.7% ARC-AGI-3 result uses a provider adapter that preserves opaque reasoning between calls and compacts long conversations. The standard harness result is lower, so the larger figure should not be treated as a model-only score unaffected by the surrounding software.

Artificial Analysis independently ranked Fable highly but reported an average 44.6 minutes and $26.54 per task in its index run. Those figures describe that demanding evaluation, not the cost of every use. They illustrate why faster token generation does not automatically translate into faster or cheaper completed work.

Anthropic also describes a 1,000-agent Lean formalization of an existing proof of Fermat’s Last Theorem, not the discovery of the theorem’s proof. Two physics manuscripts linked in the launch make narrower technical claims than a general solution to black-hole physics. Independent replication of those science results has not been established. The company’s platform documentation also says Fable normally requires 30-day retention unless Anthropic expressly authorizes an exception.

The debate is whether the narrow training limit is a useful safety measure or whether it limits legitimate research and optimization. Critics point out that improving kernels also helps people run smaller models efficiently; the size of any effect on adjacent work remains unmeasured. The system card itself could not be retrieved in the dossier’s deeper review, so detailed safety claims here rest on Anthropic’s readable launch page.

Primary source, verified: read the paper →

Ground Truth. (2026, November 10). Anthropic releases Fable and Mythos 5.2 with shared training-time limits. Ground Truth. https://groundtruth.day/news/anthropic-fable-mythos-52-capability-limits.html

@misc{groundtruth:anthropic-fable-mythos-52-capability-limits,
  title  = {Anthropic releases Fable and Mythos 5.2 with shared training-time limits},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/anthropic-fable-mythos-52-capability-limits.html}
}

Topics: models · anthropic · ai-safety