Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 in September 2026. The two are the same underlying model at different safeguard levels: Fable 5.1 is generally available, and Mythos 5.1 sits behind trusted-access programs for cybersecurity and life-sciences work. The API model ID is claude-fable-5-1.
This is a point release, not a new family. The scores behind it are the kind you usually see between whole versions, so it is worth reading closely.
The short version. General reasoning improved a few points. Autonomous, tool-using, long-running work improved a lot. Pricing held steady at $10/$50 per million tokens while cache reads got 75% cheaper. If you run agents, that combination of benchmarks and cost is the whole story.
The benchmarks: agentic work roughly doubled
The gains are lopsided by design. Everyday knowledge tasks move a little; long, multi-step tasks where the model drives a terminal or a computer move sharply.
| Benchmark | Fable 5.1 | Fable 5 |
|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% |
| AutomationBench | 31.4% | 17.1% |
| Terminal-Bench 4.0 | 55.8% | 42.0% |
| OSWorld 2.0 (partial) | 77.9% | 72.9% |
| CursorBench 3.2.0 | 73.4% | 70.5% |
| GDPval-AA v2 (ELO) | 1853 | 1723 |
| Humanity’s Last Exam (no tools) | 60.9% | 57.8% |
| Humanity’s Last Exam (with tools) | 65.0% | 63.8% |
Two lines carry the release. Terminal-Bench-Science more than doubled, and AutomationBench came close. Both measure the model finishing real work end to end, with no human in the loop. Humanity’s Last Exam, a broad reasoning test, moved only a few points by comparison. The shape is clear: this release is tuned for autonomous execution, not for trivia.
For reference, Mythos 5.1 scores 60.9% on Terminal-Bench 4.0, above Fable 5.1’s 55.8% on the same test.
The coding numbers in the system card round out the same picture. Fable 5.1 posts 81.2% on SWE-bench Pro and 89.1% on SWE-bench Multilingual, and reaches 97.5% on ARC-AGI-1 and 90.0% on ARC-AGI-2 at maximum effort. On FrontierSWE v2, where tasks run close to 20 hours each, it scores 0.57, ahead of Opus 5 (0.52) and OpenAI’s GPT-5.6 Sol (0.32). The gains cluster in exactly one place: long, autonomous coding sessions rather than quick answers.
One capability number underpins all of this: the context window is 1 million tokens (roughly 555,000 words), with a 128K-token maximum output, and the knowledge cutoff is June 2026.
Pricing: the cut lands where agents spend
Per-token pricing is $10 per million input tokens and $50 per million output tokens, unchanged from Fable 5. The change that matters for agents is cache reads, now $0.25 per million tokens, a 75% cut from the previous rate. Here is the full price list:
| Item | Price (per million tokens) |
|---|---|
| Input | $10 |
| Output | $50 |
| Cache read | $0.25 |
| Cache write (5-minute) | $12.50 |
| Cache write (1-hour) | $20 |
| Batch API input | $5 |
| Batch API output | $25 |
Cache reads at $0.25 are 2.5% of the base input rate, where every other Claude model charges 10%. The Batch API halves input and output for work that can run asynchronously. The minimum cacheable prompt is 512 tokens.
Anthropic frames the net effect as roughly 25% cheaper for typical workloads and up to 45% cheaper for highly agentic ones. That is not arbitrary. An agent replays a large, mostly stable context across many steps, so the cache read is the cost that scales with every step of a long run. Cutting it is a direct discount on autonomy. Cognition said it moved Opus 5 traffic to Fable 5.1 on launch day, because the model matched or edged Fable 5 at lower cost.
The same economics apply whether the agent runs in the cloud or on your own hardware: the expensive part of a long run is the context you replay every step. If you are weighing hosted models against a local-first setup, the cache line is where the two pricing models diverge most.
Full disclosure: we build Atomic Agent, an open-source, local-first agent runtime. If keeping the model on your own hardware matters more to you than raw frontier capability, it runs a quantized model on your own machine with no per-token bill, and takes about five minutes to try.
Agentic scientific research, with real examples
The launch leans on three worked results rather than adjectives.
Protein design. The model designed high-affinity binders with binding affinities up to 10x higher than competition winners, at a roughly 50% success rate across 12 targets where the typical rate is 10-15%.
Planetary science. It rebuilt a high-resolution Venus elevation map from 30-year-old NASA data, sharpening detail from 10-20 km down to 2-3 km and improving height accuracy by up to 25%.
Computational biology. It optimized seven deep-learning models by up to 2.5x in speed, cutting GPU cost on genome-wide analyses by 30-60% and compressing work that normally takes weeks into days.
The through line is not “knows more facts.” It is sustained, verifiable execution across long tool-using sessions, which is exactly what the benchmark jumps measure.
Two models, two safeguard levels
Fable 5.1 and Mythos 5.1 run the same weights behind different guardrails.
Fable 5.1 is the generally available version, on Claude.ai today and through AWS, Google Cloud, and Microsoft Azure. Anyone with API access can use it under the standard safeguards.
Mythos 5.1 runs with reduced safeguards for verified work, and it is gated behind two verification programs:
- Cyber Verification Program (CVP) — for defensive-security professionals. Mythos 5.1 relaxes cybersecurity safeguards so vetted defenders can use the model on work the general version would refuse. Anthropic reports no critical-severity jailbreaks found in these safeguards during testing.
- Life Sciences Verification Program (LSVP) — run in partnership with the US government, for life-sciences research professionals, with access expanding beyond the initial participants.
Mythos 5.1 is currently limited to US organizations, with international expansion under discussion. The split lets Anthropic ship a capable model to everyone while keeping the higher-risk configuration inside a vetted group rather than in the open release.
Safety and alignment
On safety, Anthropic reports improved alignment across most behavioral metrics, reduced reward-hacking versus Mythos 5, and 60% fewer false positives in the cybersecurity domain. The model is described as less likely to reach for resources outside its test environment. It can identify software vulnerabilities but is constrained against developing them.
External red-teaming came from two commissioned organizations plus Gray Swan, alongside robustness testing for prompt-injection attacks and tabletop exercises with PhD-level biologists. Anthropic says the model falls below the next risk tier in its Responsible Scaling Policy for chemical and biological threats.
Enterprise and compliance
For enterprises, Anthropic is rolling out Enterprise Frontier Safeguards (EFS) with customer-controlled data storage, plus a zero-data-retention option for eligible customers pending that rollout, across Claude Code, Enterprise, Platform, Bedrock, AWS, Google, and Azure through fall 2026.
To meet the EU AI Act, outputs carry a watermark, with a detection API in private preview for eligible organizations. Anthropic states the watermark is invisible to users and does not affect the output.
What early users report
The launch cites named organizations across coding, trading, and research. A few concrete claims stand out:
- Jane Street reports the model solves more coding problems than its predecessors and stays readable across long tasks, calling it “state-of-art on trading intuition.”
- Millennium says it found a rare crash that had gone unsolved for years, tracing it to a root cause in a vendor library.
- MongoDB built a complex prototype in three days, running the model unattended with verification loops.
- Red Hat says it identified the root cause of every broken build it tested.
- Rakuten Medical reports it found a gap other frontier models missed and proposed a new research hypothesis.
The pattern across these is the same one the benchmarks show: value comes from long, autonomous, tool-heavy runs rather than one-shot answers. It is also why a release like this raises the bar for every agent runtime, hosted or local — the reference point for “an agent can do this” just moved.
The takeaway
Fable 5.1 is a version bump that behaves like a bigger one for a specific audience. Single-turn question answering gains a few points. Agents that drive a terminal or a computer across long tasks are the real target: the doubling on Terminal-Bench-Science and AutomationBench, paired with 75% cheaper cache reads, is what moves autonomous work from demo to daily driver. That is the combination worth watching in this release.
Benchmark figures, pricing, and program details are from Anthropic’s Claude Fable 5.1 and Mythos 5.1 announcement (September 2026).