Claude Fable 5.1

Learning Objectives

• Understand what Claude Fable 5.1 is and how it relates to Mythos 5.1 and Opus 5
• Evaluate Fable 5.1's capability and benchmark claims across coding, knowledge work, and science
• Explain the safeguard-and-fallback design that makes a Mythos-class model safe to release publicly
• Read the "45 percent cheaper" claim correctly — what actually changed and what did not

What Is Claude Fable 5.1?

Claude Fable 5.1 is Anthropic's most capable publicly available model, released on September 1, 2026. It is the public, safeguarded version of Anthropic's Mythos-class frontier model — the same model line that, in its restricted form, powers Project Glasswing's cybersecurity work. The Fable 5 line launched on June 9, 2026, taking over from Claude Opus 4.8 as the top generally available Claude; the Opus line has since moved on to Claude Opus 5, which is the model Fable falls back to when a safeguard trips.

Anthropic describes the Fable line as "state-of-the-art on nearly all tested benchmarks of AI capability" across software engineering, knowledge work, vision, and scientific research. The model can work autonomously for extended periods, retaining memory across millions of tokens — Stripe reported that Fable 5 "compressed months of engineering into days" on a codebase migration.

What the 5.1 release changed

The 5.1 update is best read as three separate things, because the headline number describes only one of them.

All benchmark figures above are reported by Anthropic on its own evaluation runs, not independently verified.

Capabilities and Benchmarks

The Fable line posts the highest scores Anthropic has recorded across several independent evaluations:

Beyond coding, Fable's vision is strong enough to reconstruct working code from a screenshot and to play games without auxiliary tools. In molecular biology, Anthropic says the model outperformed specialized protein models at predicting virus shell assembly, and that domain experts preferred its novel hypotheses roughly 80 percent of the time over Opus-class models.

Safeguards and the Fallback Design

Releasing a Mythos-class model publicly required a new safety architecture. Fable runs three safety classifiers that, when triggered, fall back to Claude Opus 5 rather than answering directly. The classifiers cover:

1. Cybersecurity and offensive hacking tasks
2. Dual-use biology and chemistry research
3. Model distillation attempts (using Fable to train competing models)

Anthropic reports that roughly 95 percent of Fable sessions involve no fallback at all — the safeguards are designed to catch a narrow band of high-risk requests rather than throttle everyday use.

The biology classifier was rebuilt in August 2026

Fable 5 shipped with a biology classifier tuned deliberately toward caution, on the reasoning that a frontier biology model could give meaningful uplift to someone attempting weapons development. The cost was a large volume of false positives: ordinary health and educational questions were being pushed down to the weaker fallback model.

On August 7, 2026 Anthropic published a revision. It rewrote the classifier's constitution — the rule set defining what counts as safeguarded — carving out explicit exceptions for benign cases, collected review from internal and external experts, generated fresh training data reflecting the new rules, and retrained. Anthropic reports the update cut biology-related fallbacks by about 85 percent across its product surfaces, and says it verified the classifier still triggers on genuinely harmful content.

Fable 5.1 reports another false-positive reduction

With the 5.1 release Anthropic says the model cuts biology false positives by about 85 percent and cybersecurity false positives by about 60 percent. The cybersecurity figure is the genuinely new one: every prior tuning pass had covered biology only, leaving security work as the standing complaint.

Availability

• Claude API and consumption-based Enterprise plans: available to US persons; access for foreign nationals was restored from July 1, 2026 following the June 30 lift of export controls
• Subscription plans (Pro, Max, Team, seat-based Enterprise): included at no extra cost for up to 50 percent of weekly usage limits until July 7, 2026; since July 8, 2026 subscription access has run on usage credits at API rates. Anthropic described the shift as capacity-driven and temporary, and said it aimed to return Fable to standard subscription plans once capacity allowed; it has not announced that reversal
• Mythos 5 (safeguards lifted): restricted to Project Glasswing cybersecurity partners, expanding to biomedical researchers, with a broader trusted-access program planned

Pricing

Fable 5.1's per-token pricing ($10 input / $50 output per million tokens) sits above the standard Opus tier — the premium buys the frontier capability level. That rate is unchanged from Fable 5; what moved is the cached-read rate, down 75 percent to 25 cents per million. For routine professional work where an Opus-class model is sufficient, Opus remains the more economical default.

How Fable 5.1 Compares

Strengths

• Frontier capability, publicly available — Mythos-class performance without the invite-only gate
• State-of-the-art benchmarks — leads FrontierCode, CursorBench, and the Hebbia Finance Benchmark; 90 percent on a core analytics benchmark
• A large generational jump on agentic coding — Anthropic reports Terminal-Bench 4.0 rising from 42.0 to 55.8 percent between Fable 5 and 5.1, and Terminal-Bench-Science roughly doubling
• Long-horizon autonomy — works across millions of tokens of memory; "months of engineering into days" on real migrations
• Strong vision and science — reconstructs code from screenshots; outperformed specialized protein models on virus-shell prediction
• Safe-by-default release — three safety classifiers fall back to Opus 5 on high-risk requests, with no fallback in roughly 95 percent of sessions
• Safeguards get tuned, not just shipped — the August 2026 biology revision cut false-positive fallbacks about 85 percent while holding the dual-use line, and 5.1 extended that work to cybersecurity for the first time
• Cheaper to run agents on — cached reads fell 75 percent to 25 cents per million, which is where the "up to 45 percent" agentic saving actually comes from

Limitations and Considerations

• Over-broad safeguard triggers — the cybersecurity classifier can downgrade legitimate security and code-review work to Opus 5; approved researchers need the Cyber Verification Program. Fable 5.1 reports a 60 percent cut in these false positives, which narrows the problem rather than closing it
• Dual-use biology still routes away — virology, toxicology, and molecular design fall back to Opus 5 no matter how legitimate the request; researchers in those fields do not get frontier capability here
• The cost saving is conditional, not a rate cut — per-token pricing is unchanged, so the "up to 45 percent" figure only materializes on workloads that re-read cached context; a low-cache workload sees little of it
• Premium pricing — $10 input / $50 output per million tokens is double the standard Opus tier; not the right default for routine work
• Subscription access runs on usage credits — free on Pro, Max, and Team for up to 50 percent of weekly limits until July 7, 2026, and usage-credit based since July 8; Anthropic called this temporary and capacity-driven but has not announced a return to standard subscription access
• Export-control saga resolved — a mid-June 2026 US directive briefly suspended Fable 5 and Mythos 5 access for non-US persons after a jailbreak was found; Commerce lifted the controls on June 30 and Anthropic restored access from July 1
• Closed model — API-only; no downloadable weights
• Mythos 5 stays gated — the unrestricted capabilities remain limited to vetted partners

Related Tools

• Claude Mythos 5.1 — The same model with safeguards lifted, restricted to registered partners
• Claude Opus 5 — Fable's safeguard-fallback model and the economical default for routine work
• Claude.ai — Anthropic's flagship chat interface, where subscribers reach Fable 5.1
• Claude Code — Anthropic's terminal coding agent, a primary surface for frontier-model work

Key Takeaways

• Claude Fable 5.1 is Anthropic's public flagship, released September 1, 2026 — a safeguarded version of its Mythos-class model; the Fable line launched June 9, 2026
• Anthropic reports large generational gains on agentic coding and science: Terminal-Bench 4.0 from 42.0 to 55.8 percent, Terminal-Bench-Science from 24.7 to 52.6 percent, and Humanity's Last Exam from 57.8 to 60.9 percent without tools — all vendor-reported
• The per-token price did not change. Fable 5.1 still bills $10 input and $50 output per million tokens; the "roughly 25 percent cheaper, up to 45 percent on agentic work" claim comes from a 75 percent cut to cached reads (now 25 cents per million) and from fewer wasted retries, so a workload that does not reuse cached context will not see it
• Three safety classifiers fall back to Claude Opus 5 on cybersecurity, dual-use biology and chemistry, and model-distillation requests; roughly 95 percent of sessions see no fallback
• Fable 5.1 reports cutting cybersecurity false positives about 60 percent — the first tuning pass to address security rather than biology — and biology false positives about 85 percent, a figure Anthropic has not reconciled against the similar reduction it reported for the August 2026 classifier rewrite, so do not treat the two as compounding
• Mythos 5.1 is the same model with safeguards lifted, restricted to registered partners doing cybersecurity or life-sciences research