Anthropic Launches Claude Fable 5.1 and Mythos 5.1, Pushing AI Agents Deeper Into Coding and Science

Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1, a new frontier pair aimed at long-running coding, knowledge work and scientific research. Anthropic says the two products use the same underlying model but are deployed with different safeguards and access policies: Fable 5.1 is generally available, while Mythos 5.1 is limited to vetted organizations through trusted-access programs.

The launch matters beyond another benchmark refresh. Anthropic is positioning Fable 5.1 around work that can continue for hours, span multiple tools and recover from failures — the kind of behavior increasingly expected from production AI agents.

Quick Take

• Claude Fable 5.1 launched on September 1, 2026 and is Anthropic's new generally available frontier model for demanding reasoning and long-horizon agentic work.
• Fable 5.1 supports a 1 million-token context window and up to 128,000 output tokens.
• Standard API pricing is $10 per million input tokens and $50 per million output tokens.
• Cache reads cost $0.25 per million tokens, one quarter of Fable 5's cache-read cost.
• Anthropic says typical token-billed workloads can cost about 25% less than Fable 5, with savings of up to roughly 45% for highly agentic work.
• Claude Mythos 5.1 uses the same underlying model but remains restricted to vetted organizations for sensitive cybersecurity and life-sciences use cases.

Fable 5.1 Is Built for Long-Running Work

Anthropic describes Fable 5.1 as a model for demanding reasoning and long-horizon agentic work rather than a default choice for every task. Its official model documentation lists a 1M-token context window, 128K maximum output, always-on adaptive thinking and a June 2026 knowledge cutoff.

That positioning is important. Frontier AI competition is moving beyond single-turn answer quality toward whether a model can plan a larger job, use tools, maintain context, verify its own work and keep operating without constant human supervision.

Anthropic specifically highlights agentic coding, multistep research, document-heavy knowledge work and asynchronous tasks that can continue for hours.

Specs and Pricing

The headline input and output prices are unchanged from Fable 5. The economic change is cache reading. Anthropic cut cache-read pricing to $0.25 per million tokens, which matters particularly for agents that repeatedly reuse a large working context.

For developers, that means the effective cost of a long-running workflow can fall even when the published input and output token prices stay the same.

Anthropic Reports Large Agentic Benchmark Gains

Anthropic's own launch evaluations show the largest improvements in the areas it is emphasizing most heavily: coding, tool use, scientific research and long-horizon work.

In Anthropic's published setup, Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, compared with 24.7% for Fable 5, 29.0% for Claude Opus 5 and 22.4% for GPT-5.6 Sol.

On Terminal-Bench 4.0, Fable 5.1 scored 55.8%, while Mythos 5.1 reached 60.9%. Anthropic also reports 73.4% on CursorBench 3.2.0, 31.4% on AutomationBench and 65.0% on Humanity's Last Exam with tools.

These results should be read as vendor-reported benchmarks, not independent proof of broad real-world superiority. Benchmark harnesses, safeguards and evaluation settings can materially affect model scores. Anthropic itself provides methodological notes and uncertainty ranges for some of the comparisons.

The useful signal is the direction: the company is optimizing its frontier tier for tasks that require sustained execution rather than only strong first responses.

Mythos 5.1 Is the Restricted Version of the Same Frontier Model

Anthropic says Claude Fable 5.1 and Claude Mythos 5.1 are the same underlying model with different safeguard and access regimes.

Fable 5.1 is designed for broad availability. Mythos 5.1 is offered to vetted organizations through Anthropic's trusted-access programs, with a particular focus on advanced cybersecurity and life-sciences work.

Anthropic's Mythos materials say access is currently limited and that Mythos 5.1 starts at the same $10/$50 per-million-token pricing. The distinction is therefore not a conventional fast-versus-smart model split. It is primarily a deployment-policy split around sensitive capabilities.

That is potentially a significant precedent for frontier AI: increasingly capable models may be offered in multiple access tiers depending on the risks associated with particular domains.

Science Is Becoming Part of the Frontier-Model Race

The launch also puts unusual emphasis on scientific work.

Anthropic reports experiments in molecular design, computational biology and planetary mapping. Among the examples, the company says Mythos 5.1 produced protein-binder designs that were experimentally validated, while Fable 5.1 was used to create a higher-resolution elevation map of part of Venus from decades-old Magellan radar data.

Anthropic also reports that Mythos 5.1 optimized GPU kernels for seven open-source biology models, producing speedups of up to 2.5 times while preserving outputs.

These are company-reported demonstrations, so they should not be confused with independent evidence that autonomous AI scientists have arrived. But they illustrate a broader shift in what labs are trying to prove: not merely that a model can explain science, but that it can contribute useful work inside a scientific workflow.

Safety Is Becoming Part of Product Segmentation

More capable agentic systems create a harder safety problem because they can sustain complex workflows and use tools over longer periods.

Anthropic says Fable 5.1's newer cybersecurity safeguards generate fewer false positives than its previous generation while still restricting higher-risk activities. The company is also using trusted-access programs for Mythos 5.1 in areas where the same capabilities may be useful for legitimate research but dangerous if broadly available.

This means safety policy is no longer only an invisible layer behind the model. It is increasingly becoming part of how frontier products are packaged, priced and distributed.

AI World Scope Take

Three things stand out from this release.

First, the frontier is becoming agentic. Model makers increasingly care about whether systems can work across long, multi-step tasks rather than merely dominate static question-answer benchmarks.

Second, architecture now matters to cost. A 75% reduction in cache-read pricing can materially change the economics of persistent agents even though the headline input and output token prices did not change.

Third, access tiers may become normal for frontier capability. Fable 5.1 and Mythos 5.1 show one possible model: the same underlying intelligence deployed with different safeguards and access requirements depending on the risk profile of the work.

For buyers, the relevant question is therefore becoming less “Which model has the highest benchmark score?” and more “Which model can reliably finish the workload we care about, at an acceptable cost and risk level?”

Bottom Line

Claude Fable 5.1 is a meaningful update because Anthropic is clearly optimizing around long-running autonomous work. The 1M-token context window, 128K output ceiling and lower cache-read price are designed to make persistent agent workflows more practical.

Claude Mythos 5.1 adds another dimension: frontier capability can increasingly be separated from general availability. Anthropic is treating sensitive cybersecurity and life-sciences access as a distinct deployment tier rather than simply another public model SKU.

The model race is therefore becoming a competition not just over intelligence, but over reliability, autonomy, economics and controlled access — the ingredients that determine whether frontier AI can move from impressive demos into sustained real-world work.