Anthropic Ships Claude Fable and Mythos 5.1

Anthropic on 1 September 2026 launched Claude Fable 5.1 for general availability and Claude Mythos 5.1 for trusted-access partners, saying the pair share the same weights with two safeguard tiers. On Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6 percent against 24.7 percent for Fable 5 in Anthropic's reproduced setup, while cache reads fell to $0.25 per million tokens — a 75 percent cut that the company estimates lowers typical token bills by about 25 percent and highly agentic work by up to about 45 percent.

What Fable and Mythos 5.1 change

Anthropic describes Fable 5.1 and Mythos 5.1 as one underlying model with different safeguard postures. Fable 5.1 is the broadly available coding and knowledge-work release on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Azure under the claude-fable-5-1 identifier. Mythos 5.1 keeps more permissive cyber and life-sciences access for vetted Cyber Verification Program and Life Sciences Verification Program participants, and it also powers Claude Security scans. Base list pricing stays $10 per million input tokens and $50 per million output tokens; the headline savings come from cheaper cache reads rather than a new input/output sticker.

Anthropic's product post frames the release as a response to customer pushback on price, data retention, and over-triggering safeguards. Enterprise Frontier Safeguards will let enterprises keep monitoring data in customer-controlled clouds, with a phased fall rollout; until then, eligible customers can run Fable 5.1 with zero data retention. Cyber false-positive interventions drop by about 60 percent on average per Claude Code session relative to Fable 5, and biology safeguards fire 85 percent less often on elementary biology and medical questions — while research-grade life-sciences work still routes toward Mythos access or Opus-class models.

TechCrunch's Russell Brandom, writing the same day, noted that Mythos remains limited to registered cybersecurity and life-sciences partners and that Anthropic's system card still rates Mythos as low-risk for automated AI R&D acceleration. Related PromptCrates coverage of OpenViking's context database trending on GitHub is a reminder that long-running agent stacks are competing as hard on memory cost as on raw model scores.

Benchmarks and early scientific results

Beyond Terminal-Bench-Science, Anthropic reports Terminal-Bench 4.0 at 55.8 percent for Fable 5.1 and 60.9 percent for Mythos 5.1, CursorBench 3.2.0 at 73.4 percent, AutomationBench at 31.4 percent versus 17.1 percent for Fable 5, and Humanity's Last Exam at 60.9 percent without tools and 65.0 percent with tools. Partner quotes in the post range from Jane Street's trading-intuition praise to Cursor calling Fable 5.1 its most capable CursorBench 3.2 run. Those figures are Anthropic-reported; independent leaderboards may differ within stated noise bands.

The scientific demos are concrete. Mythos 5.1, given open-source protein design tools, reached nearly a 50 percent hit rate across 12 experimental binder targets — versus typical 10–15 percent hit rates in protein design today — and beat prior Adaptyv Bio competition winners by about 10× affinity on three named targets. Fable 5.1 produced a Creative Commons Venus digital elevation map from Magellan radar, sharpening footprint detail from roughly 10–20 kilometers to about 2–3 kilometers with heights up to 25 percent more accurate. Mythos also wrote custom GPU kernels that sped seven open-source protein and genomics models by up to 2.5× on an NVIDIA H100, with estimated genome-wide GPU cost cuts of 30–60 percent.

Safeguards, trusted access, and availability

Anthropic says Mythos 5.1 still sits below the next Responsible Scaling Policy chemical and biological risk tier and inside the lower cyber risk category of its Frontier Compliance Framework when cyber safeguards are off, even though cyber capability is the strongest it has released. Fable 5.1 may identify software vulnerabilities for defensive work but still redirects penetration testing, exploit generation, and binary-based vulnerability scanning to Opus models. Anti-distillation changes block new API accounts from manually editing prior Claude thinking context while preserving transcripts — a technique Anthropic says distillers used.

For buyers comparing frontier releases, the practical story is cheaper cache-heavy agent loops on a generally available Fable SKU, plus a gated Mythos path for cyberdefense and life-sciences R&D. PromptCrates previously tracked related containment debates around Runway's Solaris interface world model; Fable 5.1 instead sells usefulness with narrower false positives rather than a new multimodal interface. Teams should treat Mythos enrollment timelines as US-first and government-coordinated, not as a same-day global GA.

Sources

• Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic, September 2026
• Anthropic's new Fable release is cheaper, less restrictive — TechCrunch (Russell Brandom), 1 September 2026