Fable 5 and Mythos 5 Share a Model but Separate Access Rules
Anthropic launched Fable 5 for general use with Opus 4.8 fallback and Mythos 5 for restricted partners at $10/$50 per million tokens. AA reported a 64.9 launch index lead using max effort and fallback, which affected about 8% of its tasks.
Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9 as two access configurations of the same underlying model. Fable was offered for general use with classifiers and fallback routing; Mythos was restricted to vetted users with specified safeguards lifted. Both launched at $10/$50 per million input/output tokens.
Safeguards and access#
Fable routes flagged cybersecurity, biology/chemistry and distillation-related requests to Claude Opus 4.8. Anthropic said fewer than 5% of sessions triggered fallback in early data and acknowledged benign false positives. Business traffic on Mythos-class models required 30-day retention for safety, with the company saying it would not use that data for training.
Mythos 5 initially upgraded existing Glasswing access with cyber safeguards lifted. A separate biology program was planned with biology/chemistry safeguards removed while retaining cyber controls. This was not unrestricted public access.
The launch offered Fable through the API and consumption-based Enterprise. Pro, Max, Team and seat-based Enterprise subscriptions included it through June 22, with usage credits planned afterward unless capacity allowed an extension.
Claims vs checks#
Anthropic described gains in coding, knowledge work, vision and science. Its customer anecdotes and alignment assessments remain vendor-presented evidence.
Artificial Analysis, using adaptive reasoning at max effort with Opus 4.8 fallback, reported a 64.9 Intelligence Index score, first place, and 1932 GDPval-AA Elo. The configuration matters: AA observed fallback on approximately 8% of index tasks, versus Anthropic’s under-5% session statistic; the units and workloads differ. Its separate GDPval-AA report recorded 2% task fallback.
Limits#
These are launch terms and dated evaluation results, not current availability. Fallback means some measured work was answered by another model. No later Arena placement or export-control timeline is established by this launch brief.
Correction (October 14, 2026): Added AA’s evaluation effort and measured fallback rates; removed unverified retrospective Arena rankings and export-control chronology from the launch account.