On 1 September 2026, Anthropic released Claude Fable 5.1 for general availability and Claude Mythos 5.1 through its trusted-access programs. The pair is the same underlying model with different safeguard tiers: Fable 5.1 is the public frontier SKU; Mythos 5.1 relaxes cyber and life-sciences restrictions for vetted defenders and researchers.
API callers can use claude-fable-5-1 today on the Claude Platform, Amazon Bedrock, Google Cloud, and Microsoft Azure. List input and output pricing stays at $10 / $50 per million tokens, but cache reads drop 75% to $0.25 per million tokens — Anthropic estimates roughly 25% lower total cost on typical workloads and up to about 45% on highly agentic, cache-heavy jobs.
What's new
- Models: Claude Fable 5.1 (GA) and Claude Mythos 5.1 (CVP / LSVP trusted access) — same weights, different safeguards
- API ID: claude-fable-5-1
- Benchmarks (Fable 5 → 5.1): Terminal-Bench-Science 0.1 24.7% → 52.6%; Terminal-Bench 4.0 42.0% → 55.8% (Mythos 5.1 at 60.9%); OSWorld 2.0 partial 72.9% → 77.9%; Humanity's Last Exam (no tools) 57.8% → 60.9%; AutomationBench 17.1% → 31.4%; GDPval-AA v2 1723 → 1853
- Comparators (Anthropic table): Opus 5 at 29.0% on Terminal-Bench-Science; GPT-5.6 Sol at 22.4%
- Pricing: cache reads $0.25/M (−75%); input/output unchanged at $10/$50 per million tokens
- EFS: Enterprise Frontier Safeguards store customer data in customer-controlled cloud — phased enterprise rollout starting fall 2026; eligible customers can run Fable 5.1 with zero data retention until EFS ships
- Cyber safeguards: ~60% fewer false-positive interventions; Fable 5.1 may help find vulnerabilities but not develop exploits; dual-use tasks still route to Opus
- Preserved thinking: anti-distillation rule for new API accounts (documented in Help Center; existing accounts grandfathered for now)
- Mythos 5.1 access: Cyber Verification Program and new Life Sciences Verification Program (US-government partnership); Claude Security product now runs on Mythos 5.1
Why it matters
This is the GA launch the September watch guide flagged as missing after Anthropic named Fable 5.1 only in API policy docs. The headline upgrade is not a token-price cut on input/output — it is cache economics that make continuous agent harnesses materially cheaper, close enough that partners like Cognition describe moving Opus-class review traffic to Fable 5.1 on day one.
On capability, the Terminal-Bench-Science jump — more than doubling Fable 5 — is the clearest signal that Anthropic is positioning 5.1 for long-horizon research agents, not just coding. Mythos 5.1 keeps the dual-track pattern from the June Fable/Mythos 5 launch: public Fable with tighter guardrails, gated Mythos for cyber and life-sciences work.
Our take
Fable 5.1 reads as Anthropic answering three customer complaints in one release: agent cost (cache), data residency (EFS / ZDR bridge), and over-blocking (cyber false positives). The preserved-thinking API change we noted in the September AI Watch is now part of a full GA story, not a standalone policy footnote. Watch whether Mythos 5.1 CVP enrollment widens beyond the initial US cohort — that is still the bottleneck for offensive-adjacent cyber eval work.
