I spent the morning of September 3 reading Anthropic’s Fable 5.1 / Mythos 5.1 launch the way I read a model card before I move production traffic: tables first, pricing second, safeguard footnotes third. The release date on the post is September 2026; the API release notes pin general availability to September 1. For anyone running coding agents, computer-use loops, or long-horizon research harnesses, this is the week’s material model change.

What shipped

Claude Fable 5.1 and Claude Mythos 5.1 share identical weights. Fable is the generally available configuration (claude-fable-5-1 on the Claude API, plus AWS, Google Cloud, and Azure). Mythos is the same checkpoint with looser domain safeguards for vetted cyber and life-sciences programs. The system card lists a June 2026 knowledge cutoff. Platform notes specify a 1M-token context window by default, 128k max output tokens, and always-on adaptive thinking.

Anthropic did not publish parameter counts. Treat size as undisclosed.

The numbers that matter for routing

Anthropic’s comparison table is the useful artifact. On Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6% versus Fable 5 at 24.7%, Opus 5 at 29.0%, and GPT-5.6 Sol at 22.4%. That is not a one-point polish pass; it is a step change on a science-agent harness with stated standard error of roughly ±3.5–4.5 points.

On Terminal-Bench 4.0, Fable 5.1 lands at 55.8% and Mythos 5.1 at 60.9%, against Fable 5 42.0%, Opus 5 52.3%, and Sol 37.3%. Anthropic attributes the Fable–Mythos gap partly to cyber-safeguard interventions on Fable; they say today’s safeguard updates should shrink that gap. If you evaluate with production Fable policies on, expect the lower number.

CursorBench 3.2.0 is 73.4% for Fable 5.1 versus 70.5% / 70.0% / 67.2% for Fable 5, Opus 5, and Sol. Cursor’s quote on the launch page calls it the most capable model they have run on that suite at max effort. Absolute gains here are smaller than on Terminal-Bench-Science, which matches what I see when a lab is already near the top of a coding harness: the interesting move is consistency and verification, not a 20-point leap.

Other named scores from the same table: GDPval-AA v2 1853 (Fable 5 1723, Opus 5 1824, Sol 1711); AutomationBench 31.4% (Fable 5 17.1%, Opus 5 26.9%, Sol 19.6%); Humanity’s Last Exam 60.9% without tools and 65.0% with tools; OSWorld 2.0 77.9% partial / 41.7% strict on the authors’ August 2026 task release. Anthropic warns those OSWorld numbers are not comparable to older published OSWorld 2.0 figures because the task files changed.

Read the safeguard footnote carefully. On tasks where Fable’s production safeguards intervened, Anthropic scored zeros and sometimes fell back to Opus 4.8 (cyber) or Opus 5 (biology). That understates Fable relative to an unrestricted competitor on those suites. If your eval includes dual-use cyber or research-bio prompts, split the scorecard: “with Fable production policy” versus “Mythos / trusted access.”

Price is the other half of the routing decision

List price for Fable 5.1 matches Fable 5: $10 per million input tokens and $50 per million output tokens. The operational change is cache reads at $0.25 per million tokens — Anthropic’s stated 75% cut. They estimate ~25% lower spend on typical Enterprise / Claude Code / API mixes measured over four weeks of August 2026 usage, and up to ~45% on highly agentic, cache-heavy workloads.

For context, OpenAI’s GPT-5.6 Sol API page currently lists $4 / $20 per million input/output (promotional pricing through at least November 21, 2026), with a 1,050,000-token context window and 128,000 max output. On raw list rates Fable remains 2.5× Sol on both input and output. The only way Fable wins a pure cost bake-off is if cache hit rates are high and/or you need fewer tokens per successful task. Cognition’s note on the Anthropic post — moving Devin’s Opus 5 traffic to Fable 5.1 because cost per task fell once cache reads dropped — is exactly that calculation.

Defaults matter for metering: Anthropic says Fable 5.1 defaults to High effort in Claude Code and Medium in Claude Cowork / claude.ai. Low/Medium effort is where they claim quality similar to old High at lower cost. Measure your harness at the effort you will actually ship.

Where it breaks

I am not treating the science demos as production SLAs, but they are part of the launch claim set: Mythos-assisted binder designs with reported hit rates near 50% across 12 targets and binding affinities up to 10× better than prior Adaptyv Bio competition winners on three named targets; a Venus elevation map released under Creative Commons; and up to 2.5× inference speedups on seven open protein/genomics models via custom GPU kernels (NVIDIA H100 figures on the post), with estimated 30–60% GPU cost cuts on genome-wide jobs. Treat those as lab results with Anthropic’s footnotes, not as your cluster’s guaranteed FLOPS.

Limits Anthropic itself flags:

  • Cyber: Mythos is their strongest cyber model yet, still in the lower risk category of their Frontier Compliance Framework. Fable may help discover vulnerabilities but still redirects exploit development, pen testing, and related dual-use work.
  • Biology: research-grade capabilities stay gated; Mythos access via a Life Sciences Verification Program built with the US government.
  • Alignment: better than Mythos 5 on several automated behavioral metrics, but the model can still bypass approvals and auto-mode classifiers; coverage of very long multi-agent and “impossible task” settings is incomplete.
  • Distillation hardening: new API accounts lose the ability to manually edit prior thinking while keeping the thinking transcript — a breaking change for some custom multi-turn tooling.
  • Compliance: EU AI Act watermarking on post–August 2, 2026 models; detection API in private preview.
  • Enterprise Frontier Safeguards (customer-held storage with misuse detection) roll out in phases starting later this fall; eligible customers get zero data retention on Fable 5.1 until then.

Param count, training FLOPs, and independent third-party replications of the new Terminal-Bench-Science and Terminal-Bench 4.0 numbers were not in the materials I pulled this morning. Until those land, I treat Anthropic’s table as vendor numbers with disclosed harness notes — useful for ranking, not as gospel.

How I would actually migrate

If I already run Fable 5 or Opus 5 for coding agents, I would A/B Fable 5.1 on the same harness with cache metrics enabled, at Medium and High effort, and compare cost per resolved task, not cost per token. If my traffic is Sol-priced and cache-light, Sol’s $4/$20 still wins on the invoice until Fable’s quality-per-token closes the gap on my tasks. If I need Mythos-class cyber or bio research behavior, I am in an access program, not on the public Fable endpoint.

Same week, NVIDIA published production notes for Groq 3 LPX, a Vera CPU shipping update, NVHBM, and GB300 NVL72 InferenceX economics. Those matter if you buy racks. For the people who buy model endpoints and measure agent pass rates, Fable 5.1 is the story worth leading with this morning.

Anpoo Sivanadi, Staff Software Engineer