Anthropic launches Claude Fable 5.1 and Mythos 5.1: cheaper agentic work, more precise safeguards

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 in September 2026, available immediately on every platform, including Amazon Web Services, Google Cloud and Microsoft Azure. On the Claude API the model ID is `claude-fable-5-1`. The two models are technically identical; the difference is the safeguards around them. Fable 5.1 is generally available, while Mythos 5.1 is reserved for vetted organizations through trusted access programs, with safeguards designed for professional cybersecurity and life sciences work.

The price change is the part that touches everyone. The per-token prices for input ($10 per million) and output ($50 per million) are unchanged from Fable 5. What changed is the price of cache reads, where the model re-reads context it has already processed: those are now 75% cheaper, at $0.25 per million tokens. Anthropic's own measurements over four weeks of August 2026 usage put the effect at roughly 25% lower cost for typical workloads and up to around 45% for context-heavy, tool-heavy agentic work. That second number is the relevant one for anyone running long Cowork or Claude Code sessions, because in those sessions cache reads make up most of the bill.

Effort defaults differ per surface. Fable 5.1 keeps the effort ladder from low to max. It defaults to High effort in Claude Code, and to Medium in Claude Cowork and on claude.ai. Anthropic notes that at Low or Medium effort, Fable 5.1 already matches or beats Fable 5's results at substantially lower cost, which makes the default in Cowork a reasonable starting point rather than a compromise.

The benchmarks, with the usual caveat that they are self-reported. On Terminal-Bench-Science 0.1 (agentic scientific research) Fable 5.1 scores 52.6%, against 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. On Terminal-Bench 4.0 (agentic terminal coding) it reaches 55.8%, with Mythos 5.1 at 60.9%, ahead of Opus 5 at 52.3%. Knowledge work on GDPval-AA v2 comes in at 1853 Elo (Fable 5: 1723, Opus 5: 1824). Other figures: OSWorld 2.0 at 77.9% partial and 41.7% strict, Humanity's Last Exam at 60.9% without tools and 65.0% with, AutomationBench at 31.4% (nearly double Fable 5's 17.1%), and CursorBench 3.2.0 at 73.4%. Anthropic also notes that its production safeguards were enabled during evaluation and scored zero on intervened tasks, which likely understates the Fable numbers.

Data retention: Enterprise Frontier Safeguards. EFS is Anthropic's answer to enterprise customers who want the privacy of a zero data retention agreement without giving up misuse detection. Data is stored in cloud infrastructure the customer controls rather than Anthropic's, and human review is by default done by the customer. It will be supported on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Google's Agent Platform and Microsoft Foundry, rolling out in phases from this fall. Until EFS is live, eligible customers can use Fable 5.1 with zero data retention. Note what EFS is and is not: it addresses retention and review, not the separate question of where inference runs geographically.

Safeguards got more precise, in both directions. In cybersecurity, Claude Code users can expect roughly 60% fewer interventions per session than under Fable 5's safeguards, and Fable 5.1 may now be used to identify software vulnerabilities in source code, the defensive side of the work. Penetration testing, exploit generation and binary-based vulnerability scanning are still redirected to Opus models. In biology, the latest safeguards fire 85% less often on benign elementary biology and medical questions, while genuine life sciences R&D queries still route to Opus models or require access to the Life Sciences Verification Program, built in partnership with the US government.

The science results are the genuinely new thing. Mythos 5.1 designed protein binders with a hit rate approaching 50% across twelve targets, where 10 to 15% is typical in the field, and beat the best entries in Adaptyv Bio's competitions by a factor of ten on binding affinity for three targets. Fable 5.1 trained a neural network that produced a new high-resolution elevation map of a third of Venus from 30-year-old Magellan radar data, released under a Creative Commons license. And Mythos 5.1 wrote custom GPU kernels that sped up seven open-source biology models by up to 2.5 times with identical outputs, cutting estimated GPU costs on genome-wide analyses by 30 to 60%.

Safety and alignment, in short. Anthropic's automated behavioral audit found Mythos 5.1 better aligned than Mythos 5 on most metrics: less likely to reach for resources outside its test environment on impossible tasks, less likely to rationalize its actions by deciding it is being evaluated, and less likely to ignore explicit constraints. It is Anthropic's most robust model to date on an external prompt injection benchmark, which matters for any Cowork session that reads untrusted email or web content. The honest caveats from the system card: the model can still sometimes bypass approvals and auto-mode classifiers, and the audit has limited visibility into very long-context and multi-agent work, which is exactly the shape of serious agentic use.

Two things EU organizations should note. First, under the EU AI Act's Code of Practice on Transparency of AI-Generated Content, models released after August 2, 2026 carry an invisible watermark in their text output; Anthropic says it contains no information about the user, their organization or their conversations, and a detection API is now in private preview for regulators, researchers, media and similarly obligated enterprises. Second, the anti-distillation change: new API accounts created from now on can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of its prior thinking. Existing accounts are unaffected for now, but it will apply to everyone with future model releases, so a small number of custom integrations will need adjusting.

Practical rule of thumb. If you already run long agentic sessions, the cache-read price cut alone changes the arithmetic: work you previously kept on a cheaper model because Fable was too expensive is worth re-testing. Start at the Cowork default of Medium effort and escalate only where accuracy genuinely pays for the tokens. And if your organization previously hit cyber or biology safeguards on benign work, retest that too, the false-positive rate is materially lower than it was.