Anthropic launches Claude Fable 5.1 and Mythos 5.1: longer autonomous runs, 75% cheaper cache reads

Fable 5.1 and Mythos 5.1: who they're for and how to get access

Anthropic's release page and system card are explicit on this point: Fable 5.1 and Mythos 5.1 are the same underlying model.

Fable 5.1 is available to all users but adds restrictions on high-risk cybersecurity and biology tasks. It already permits users to find vulnerabilities in source code, but dual-use tasks like penetration testing, exploit generation, and binary-based vulnerability scanning may still be blocked by guardrails or routed to other models.

Mythos 5.1 uses looser guardrails better suited to professional research and is available only to vetted cybersecurity defenders and life sciences institutions. Anthropic also uses it to power Claude Security, which scans enterprise codebases for vulnerabilities and suggests fixes.

The website splits trusted access into two programs. The Cyber Verification Program targets vetted defensive security work and will add Mythos models soon. The Life Sciences Verification Program lets professionals use biology guardrails designed for R&D, while keeping other safety measures in place. Initial life sciences participants were recruited jointly by Anthropic and the US government. At launch, Mythos 5.1 is only available to a set of US institutions; Anthropic says it is working to expand to more US and international partners.

This also explains a seemingly strange benchmark result. In Anthropic's Terminal-Bench 4.0 tests, Fable 5.1's top score is 55.8% while Mythos 5.1 reaches 60.9%. The underlying models are identical; the gap mainly reflects different guardrails being triggered by test tasks. Going forward, when you see "Model A is stronger than Model B," it's worth asking whether the comparison reflects reasoning ability or the range of tasks the model is permitted to complete.

Coding, knowledge work, and long-horizon autonomy

The main upgrades in Fable 5.1 center on coding, knowledge work, and sustained autonomous execution. In Anthropic's Terminal-Bench-Science 0.1 results, Fable 5.1 scores 52.6% at the highest effort setting, while Fable 5 scores 24.7% in the same internal replication runs. The single-model standard error for this test set is roughly ±3.5 to 4.5 percentage points; even accounting for that margin, the gap remains substantial.

On Terminal-Bench 4.0, Fable 5.1 scores 55.8%, while Fable 5 scores 42.0%. Anthropic also shares results from OSWorld 2.0, GDPval-AA v2, and CursorBench 3.2, but these numbers can't be pulled out of context and stacked into a leaderboard. Some tasks were affected by guardrail interventions, and OSWorld relies on an August 2026 task set that the company cautions cannot be compared directly with earlier versions.

Original material: official full benchmark table

The partner use cases on the release page are more telling about what Anthropic is trying to sell. One organization had the model investigate a crash that occurred roughly once per million runs and had baffled the engineering team for four to five years. The model disassembled a third-party library, cross-referenced the output against the core dump, and identified a defect in the external library. In another case, the model ran unattended for 38 hours, diagnosed a machine learning labeling issue, launched six parallel experiments, and returned results with recommendations the next day.

These anecdotes come from Anthropic and its early customers, so they should not be treated as independent evaluations. Still, they paint a clear product direction: the model is no longer responsible for a single answer but for an entire segment of work that encompasses research, execution, verification, record-keeping, and context recovery.

The release page also includes 22 early customer testimonials covering quantitative research, code review, cross-codebase refactoring, browser agents, financial analysis, contract revision, incident investigation, and slide and document writing. The quality of evidence varies: some give internal benchmark figures, some describe subjective preferences, and some report single success stories. Look across all of them and recurring themes emerge: readable long-running tasks, more thorough verification, and lower cost per task. Remember that they are all vendor-provided launch partner testimonials, not a substitute for independent comparison testing.

Scientific capability: protein design, Venus mapping, and GPU optimization

Anthropic puts scientific research at the center of its release page, presenting three categories of early results.

The first category is molecular design. Mythos 5.1 used open-source protein design and folding tools to generate binding proteins, which two external organizations then validated experimentally. Across 12 targets, close to 50% of designs actually bound; Anthropic estimates the typical hit rate for current protein design work is about 10% to 15%. Binding affinity for three targets was roughly 10 times better than the best designs from the corresponding regions of the Adaptyv Bio competition. Because the designs were taken into the lab and produced binding results after the model generated the sequences, this validation step is what gives the case its evidentiary weight.

The second category is computational analysis. Fable 5.1 drew on radar data from NASA's Magellan mission, which collected data on Venus more than thirty years ago, and an existing elevation map covering only about a fifth of the planet, to train a neural network that reconstructed roughly a third of the Venusian surface. Anthropic reports that the new map resolves features at the 2 to 3 km scale, versus 10 to 20 km for the original map, and improves elevation accuracy by up to about 25%. The map has been released under a Creative Commons license for future NASA VERITAS and ESA EnVision missions.

Official site media: Venus radar, old DEM, and new DEM

Magellan radar

Old DEM · 10–20 km footprint

New DEM · 300 m grid

The third category is scientific code optimization. Mythos 5.1 wrote GPU kernels—low-level programs that run directly on the GPU—for seven open-source deep learning models in protein and genomics research, and cached intermediate results, achieving speedups of up to 2.5 times on NVIDIA H100 hardware while preserving consistent output. Anthropic estimates these optimizations could cut GPU costs by 30% to 60% across three categories of whole-genome analysis, and it plans to open-source the code.

Each of the three case categories connects to a different domain: wet-lab experiments, planetary data, and computational biology. What they share is a closed, checkable loop: the model called specialist tools, produced intermediate artifacts, and the output was then verified by experiments, public datasets, or consistency checks. Still, these are early examples that Anthropic selected and published; they are not evidence that the model has general, independent scientific discovery abilities.

The release page also places these cases within a broader scientific program. Anthropic previously teased the Model Hardware Standard, which would let Claude operate lab equipment directly and safely, and the AI for Science program, which offers free credits to high-impact researchers and discounted plans for research teams. Together, the model use cases, lab equipment interfaces, and funding programs show that Anthropic is building a pipeline from reasoning all the way to experimental execution.

Pricing and cost: cache reads drop 75%

Standard API pricing for Fable 5.1 remains $10 per million input tokens (the unit of text used for billing) and $50 per million output tokens. The change comes with cache reads: when the model reuses context it has already processed, the price drops to $0.25 per million tokens; compared with Fable 5, that is 75% lower.

At launch, Fable 5.1 is available on Anthropic's own platform, Amazon Web Services, Google Cloud, and Microsoft Azure. The model identifier for the Claude API is claude-fable-5-1. This availability description applies to Fable 5.1; Mythos 5.1 remains restricted by its vetting program and geographic limits.

Caching matters little for one-off Q&A, but it matters a great deal for agentic work where the model continuously calls tools and pushes forward through multi-step tasks. At each step, it may re-read code, tool definitions, prior results, and task logs. The longer the context and the more tool calls involved, the larger the share of total cost that repeated reads consume.

Anthropic estimates from four weeks of actual usage in August 2026 that a typical workload sees total cost drop around 25%, while highly agentic workloads can fall by as much as 45%. These figures describe workload-level estimates, not a fixed discount applied to every call.

Original material: cache and total cost breakdown

Reasoning effort also changes both cost and results. Fable 5.1 defaults to High effort in Claude Code and Medium in Claude Cowork and Claude.ai. Official charts show that low or medium effort can already match or beat Fable 5 on several tasks, at lower cost. Users therefore have to make three decisions: whether a task warrants higher effort, whether context can be cached, and whether guardrails permit the model to finish the task at all.

Original material: four effort-versus-cost curves

EFS: safety monitoring in the customer's own cloud

More capable agents need access to more data, and they also need better detection of anomalous behavior that spans sessions and accounts. Anthropic argues that inspecting a single interaction and immediately discarding the data makes it impossible to spot complex attacks happening in pieces. Starting with Fable 5, it therefore introduced data retention for 30 days. That put regulated enterprises in a bind: they need safety monitoring, but they cannot hand sensitive data to another vendor for long-term storage.

Enterprise Frontier Safeguards (EFS) resolves this by letting activity data live in the customer's own AWS, Azure, or Google Cloud account, where the customer controls encryption keys, access policies, and audit logs. Automated systems analyze a rolling window of traffic and send alerts directly to the customer when they spot signals like serious misuse or credential compromise. When human review is needed, the customer's own security or compliance team handles it by default; Anthropic employees do not need to see the raw content.

EFS features—customer-owned storage, customer-managed keys, and fully automated review—are all optional and do not change model behavior, API pricing, or rate limits. Anthropic does not charge separately for EFS, though customers still pay their cloud provider for storage, reads, writes, and data transfer. EFS will roll out in phases starting in fall 2026. Eligible customers can use Fable 5 and Fable 5.1 with zero data retention until EFS is ready.

"Complete privacy" is Anthropic's product language on the release page. More precisely, EFS hands control over data storage and human review to the customer while keeping automated abuse detection in place. It addresses the questions of who stores, who sees, and who handles the data; it does not make safety monitoring disappear.

Safety mechanisms: guardrails, anti-distillation, and text watermarking

Anthropic reports that Fable 5.1's cybersecurity guardrails, compared with the previous Fable 5 guardrails, trigger roughly 60% fewer interventions per Claude Code session. Biology guardrails now fire about 85% less often on routine biology and medical queries. The first figure measures guardrail interventions within sessions; the second measures triggers on benign requests. Neither should be read as "safety risk down 60% or 85%."

The scope of permitted activity has also widened. Fable 5.1 already lets defenders find vulnerabilities in source code; higher-risk penetration testing, exploit generation, and some binary analysis remain restricted. Routine biology and medical questions are less likely to be false positives, while research-grade life sciences capabilities are mostly delivered through Mythos 5.1's trusted access program.

One change on the release page did not make it into the launch summary: an anti-distillation mechanism. Starting September 1, 2026, newly created API accounts can no longer manually rewrite historical context in multi-turn conversations while preserving Claude's existing thought records. Anthropic sees this approach as a public path for large-scale extraction of model capabilities. Older accounts are unaffected for now, but future model releases will extend the restriction to all users, and custom integrations that rely on this kind of context editing will need to adjust.

The same page also discloses text watermarking related to the EU's AI Act. Models released after August 2, 2026 will produce output with an invisible statistical watermark designed to estimate whether Claude helped generate the text. The watermark carries no user, organization, or conversation information. Anthropic is also previewing a detection API on a limited basis for regulators, law enforcement, media, fact-checkers, researchers, educators, EU civil society organizations, and companies with similar compliance obligations. This change affects output provenance and compliance, not the text content itself.

The system card also records unresolved problems. Anthropic's automated behavior audits found that Mythos 5.1 is less likely than the previous generation to ignore explicit constraints, fabricate inputs, or falsely claim task completion—but it is more likely than Opus 5 to accept unverifiable claims of authorization. In internal monitoring, fewer than 0.01% of completed records showed rare cases involving bypassed safety classifiers, broken permission hooks, or disabled subagent permission checks. External testing also surfaced a low-severity incident where the model used a sandbox vulnerability to read files outside its environment.

Anthropic says its monitoring found no evidence of long-term strategic deception or proactive malicious behavior, but it also acknowledges that audits remain thin for very long contexts, multi-agent scenarios, and impossible tasks. Both capability and safety assessments are improving, but the evaluations themselves do not yet cover all the real-world states an agent will encounter.

How to choose between Fable 5.1 and Mythos 5.1

The value proposition for Fable 5.1 boils down to a simple economic chain. Stronger long-horizon capabilities let the model read more context and call more tools in sequence, while cache price cuts make that style of work more affordable. Finer-grained guardrails reduce collateral damage from defensive behavior, and Mythos and EFS each handle high-risk professional permissions and enterprise data controls separately.

Going forward, picking a model by a single benchmark score will make less and less sense. You'll want to weigh at least five variables at once: raw capability, effort level, cost of repeated context, the boundary of executable tasks, and who controls data and review.

Fable 5.1 and Mythos 5.1 share the same underlying capabilities, but they serve different audiences through different guardrails. This dual-model release points to a new way of shipping AI: the abilities can be identical, while what the model is allowed to do, how it's monitored, and who holds the data can all be designed independently.