Intel / article

Claude Fable 5 Mythos

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Jun 9, 2026

Context

Anthropic's Claude Fable 5 is, in Artificial Analysis's words, the first publicly available Mythos-class model (artificialanalysis.ai). It shares its underlying model with Claude Mythos 5 but adds for potentially harmful queries about cybersecurity, biology, chemistry and . When a guardrail flags a request, a new fallback mechanism can send it to a second model, such as Claude Opus 4.8, instead of answering with Fable 5 itself.

Anthropic gave Artificial Analysis early access, and Fable 5 took first place on GDPval-AA with 1932, putting Anthropic models in three of the top four places. GDPval-AA is the firm's of agentic real-world work, which it describes elsewhere as covering 44 occupations across 9 industries (artificialanalysis.ai). The run used adaptive reasoning at maximum effort with Opus 4.8 set as the fallback.

That setup matters for reading the score: it measures Fable 5 together with its fallback, not Fable 5 alone. Here the effect was small, because fallback happened on 2% of tasks, against Anthropic's statement that it occurs in fewer than 5% of sessions on average. The rate depends on the subject matter. In the same firm's Intelligence Index testing, fallback reached about 8% of tasks, mostly science questions. In August, Anthropic said a retrained biology classifier cut biology-related fallbacks by about 85% in its testing, while dual-use requests in areas such as virology and toxicology still fall back, now to Opus 5 (anthropic.com).

Key takeaways · AI-distilled

Per Artificial Analysis, Claude Fable 5 shares the same underlying model as Claude Mythos 5, with added guardrails for cybersecurity, biology, chemistry and distillation-related queries.
Its 1932 GDPval-AA score was measured with adaptive reasoning at max effort and Opus 4.8 configured as the fallback model, so the ranking reflects that combined setup.
Artificial Analysis observed fallback on 2% of GDPval-AA tasks, while Anthropic states fallback happens in fewer than 5% of sessions on average.
With Fable 5 at #1, Anthropic models held three of the top four GDPval-AA positions at the time of the June 2026 write-up.

Terms in this piece · Glossary

guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source artificialanalysis.ai