Claude Fable 5.1

Claude Fable 5.1

Quick answer: Claude Fable 5.1 is Anthropic's September 2026 update to the Fable line, generally available on all major clouds (AWS, Google Cloud, Microsoft Azure) via the API as claude-fable-5-1. It scores 52.6% on Terminal-Bench-Science 0.1, 55.8% on Terminal-Bench 4.0, a 1853 Elo on GDPval-AA v2, and 73.4% on CursorBench — sizeable gains over its predecessor, Fable 5, at roughly 25% lower typical cost (up to ~45% lower for highly agentic workloads) thanks to a 75%-cheaper cache-read price.

Access Status

Unlike Claude Fable 5 (suspended under a June 2026 US export control directive), Claude Fable 5.1 is generally available today across the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Azure. Its higher-safeguard sibling, Claude Mythos 5.1, is restricted to Anthropic's Cyber Verification Program (CVP) and Life Sciences Verification Program (LSVP).

At a Glance

Where Claude Fable 5.1 leads

• 55.8% Terminal-Bench 4.0 — up from Fable 5's 42.0% on the same suite
• 52.6% Terminal-Bench-Science 0.1 — more than double Fable 5's reproduced 24.7%
• 1853 Elo on GDPval-AA v2 — ahead of Claude Opus 5 (1824) and GPT-5.6 Sol (1711)
• 73.4% CursorBench — new reported high, ahead of Grok 4.6's 69.9%
• 31.4% AutomationBench, versus 17.1% for Fable 5
• ~25% lower cost than Fable 5 for typical workloads; up to ~45% lower for highly agentic, cache-heavy workloads

Where it lags

• Pricing unchanged from Fable 5 at the token-rate level: $10 / $50 per 1M input / output tokens (savings come from cheaper cache reads, not headline rates)
• Cybersecurity and life-sciences dual-use tasks are still redirected to Opus-class models or require Mythos 5.1 trusted access
• Evaluated with production safeguards enabled, which Anthropic notes can suppress scores on OSWorld 2.0 and AutomationBench when they intervene

Best for: Coding agents, long-running agentic workflows, and knowledge-work tasks where both accuracy and per-task cost matter.

What Claude Fable 5.1 Is

Claude Fable 5.1 is Anthropic's incremental update to the Fable line, announced September 3, 2026 alongside Claude Mythos 5.1. Anthropic describes Fable 5.1 and Mythos 5.1 as the same underlying model shipped with different safeguard configurations — Fable 5.1 for general availability, Mythos 5.1 for vetted cyberdefense and life-sciences use through trusted access programs.

Anthropic frames the release around three customer-feedback areas: price (an estimated 25% reduction for typical workloads, driven by a 75% cut to cache-read pricing to $0.25 per million tokens), data retention (new Enterprise Frontier Safeguards letting customers store data on their own cloud infrastructure), and safeguards precision (cybersecurity false positives cut by roughly 60%, with Fable 5.1 now permitted to identify — but not exploit — software vulnerabilities).

On benchmarks, Fable 5.1 shows broad gains over Fable 5 across agentic scientific research, agentic coding, knowledge work, computer use, multidisciplinary reasoning, business-workflow automation, and in-IDE coding tasks, while running at default "High" effort in Claude Code and "Medium" effort in Claude Cowork and Claude.ai.

Specifications

Pricing

Claude Fable 5.1's per-token pricing is unchanged from Fable 5: $10 per 1M input tokens / $50 per 1M output tokens. The cost reduction instead comes from cache reads, cut 75% to $0.25 per million tokens — Anthropic reports this yields ~25% lower cost for typical workloads and up to ~45% lower cost for highly agentic, context-heavy work where cache reads dominate spend.

Benchmark Results

*Fable 5, Opus 5, and GPT-5.6 Sol figures on this specific comparison table are Anthropic's own reproduction for the Fable 5.1 launch and may differ slightly from previously published scores for those models on Benchgen due to methodology/harness differences (e.g. Fable 5's separately reported GDPval-AA v2 score of 1760).

On Terminal-Bench 4.0, Anthropic also reports a 60.9% score for Claude Mythos 5.1 (Fable 5.1's higher-safeguard, trusted-access-only sibling) — the strongest score across all evaluated models.

Scores reported by Anthropic, September 2026, evaluated with production safeguards enabled. Anthropic notes safeguard interventions can reduce Fable 5.1's/Fable 5's scores on OSWorld 2.0 and AutomationBench, since cybersecurity and biology tasks that trigger interventions are completed by Claude Opus models instead.

Frequently Asked Questions