Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, its most powerful AI models to date. Announced on September 1, 2026, the two models share identical underlying weights but differ in their safeguard configurations. Fable 5.1 is generally available to all users, while Mythos 5.1 remains restricted to vetted organizations working in cybersecurity and life sciences. Together, they represent a significant leap in agentic coding, knowledge work, and scientific research capabilities, while introducing a novel enterprise privacy framework and substantial cost reductions.
What Are Claude Fable 5.1 and Mythos 5.1?
Claude Fable 5.1 and Claude Mythos 5.1 are the same model running behind different safety layers. Fable 5.1, available as claude-fable-5-1 on the Claude API, carries production-grade safeguards that filter requests related to cybersecurity exploits and life sciences research. Mythos 5.1 removes those restrictions for vetted participants in Anthropic’s trusted access programs, enabling deeper work in defensive security and biological research.
Both models support a 1-million-token context window, 128,000-token maximum output, and always-on adaptive thinking. They succeed the Fable 5 and Mythos 5 line that Anthropic first shipped on June 9, 2026, which were temporarily suspended later that month due to U.S. export control actions tied to cybersecurity concerns before access was restored by the end of June.
Fable 5.1 defaults to High effort in Claude Code and Medium effort in Claude Cowork and on Claude.ai. Users can adjust effort levels per message, a new beta feature that lets developers dial up or down how hard the model works on a given step without starting a new session.
Benchmark Performance: A New Standard for AI Models
Claude Fable 5.1 sets new marks across multiple industry benchmarks, often doubling the performance of its predecessor. The table below summarizes the key results Anthropic reported using its own evaluation harness:
| Benchmark | Fable 5.1 | Mythos 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench-Science 0.1 (agentic research) | 52.6% | — | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 (agentic coding) | 55.8% | 60.9% | 42.0% | 52.3% | 37.3% |
| CursorBench 3.2.0 (agentic coding) | 73.4% | — | 70.5% | 70.0% | 67.2% |
| Humanity’s Last Exam (no tools) | 60.9% | — | 57.8% | 56.6% | — |
| Humanity’s Last Exam (with tools) | 65.0% | — | 63.8% | 63.6% | — |
| AutomationBench (business workflows) | 31.4% | — | 17.1% | 26.9% | 19.6% |
| OSWorld 2.0 strict (computer use) | 41.7% | — | 36.1% | 39.6% | — |
| GDPval-AA v2 (knowledge work) | 1853 | — | 1723 | 1824 | 1711 |
The most striking result is on Terminal-Bench-Science 0.1, where Fable 5.1 scores 52.6%, more than double Fable 5’s 24.7%. Anthropic reports a standard error of 3.5 to 4.5 points per model on this benchmark.
On Terminal-Bench 4.0, the gap between Fable 5.1 (55.8%) and Mythos 5.1 (60.9%) reflects the performance cost of Fable 5.1’s stricter safeguard interventions, not a capability difference between the two models.
Early access partners confirmed these gains in real-world testing. Millennium, an investment firm, reported that Fable 5.1 identified the root cause of a rare crash in its internal systems that no engineer or previous model had explained in four to five years. Jane Street Capital noted that Fable 5.1 “remains readable over long, multi-step tasks,” while Cognition said it would move all Opus 5 traffic in its Devin product to Fable 5.1 on launch day.
Pricing and the 75% Cache Read Discount
Base pricing for Fable 5.1 remains unchanged from Fable 5 at $10 per million input tokens and $50 per million output tokens. The major change is a 75% reduction in cache read costs, from $1.00 to $0.25 per million tokens. This is the single biggest lever for cost reduction, since agentic workloads that maintain long contexts rely heavily on cached inputs.
Anthropic estimates that typical workloads will see roughly 25% lower costs compared to Fable 5. For comparison, Meta’s rival release offers Opus-level coding at contributor prices from $0.10 per million tokens. For highly agentic tasks with heavy context reuse, such as long-running coding sessions or multi-step research workflows, savings can reach approximately 45%. Batch processing is available at $5 per million input tokens and $25 per million output tokens.
This pricing strategy positions Fable 5.1 as what Every CEO Dan Shipper described as “Fable-level intelligence, Opus-level price, Sonnet-speed.” In Every’s testing, Fable 5.1 was about twice as fast as Opus 5 and used half as many tokens, making it an economical upgrade for teams previously relying on Opus for daily work.
Scientific Breakthroughs: From Protein Design to Venus Mapping
Beyond coding and knowledge work, Anthropic tested Fable 5.1 and Mythos 5.1 across scientific research domains with results that suggest AI models are approaching the ability to make meaningful contributions to discovery.
Protein Binder Design
When given access to open-source protein design and folding tools, Claude Mythos 5.1 designed high-affinity binders that were experimentally validated by two external organizations. On three targets (EGFR, Nipah G, and 15-PGDH), its binding affinities were 10 times higher than the best designs submitted to Adaptyv Bio’s protein design competitions. The model achieved a hit rate of nearly 50% across 12 targets, compared to the typical 10-15% seen in protein design today. High-affinity binders are a critical first step in developing many common drug modalities, making this a significant result for pharmaceutical research.
Venus Elevation Mapping
Claude Fable 5.1 trained a neural network to produce a new, high-resolution elevation map of roughly one-third of Venus using radar images from NASA’s Magellan mission, which collected data more than 30 years ago. The new map reveals surface details down to 2 to 3 kilometers, compared to the 10 to 20 kilometer resolution of previous maps, and shows heights up to 25% more accurately. Anthropic is releasing this map under a Creative Commons license ahead of the upcoming NASA VERITAS and ESA EnVision missions to Venus.
GPU Kernel Optimization for Computational Biology
Mythos 5.1 wrote custom GPU kernels and caching strategies that sped up seven open-source deep learning models used in computational biology by up to 2.5 times, with identical outputs. These optimizations cut estimated GPU costs by 30-60% on genome-wide analyses. Tasks that would normally take a team of performance engineers weeks to complete, and are often unaffordable for academic labs, were accomplished by Mythos 5.1 in days using publicly available source code alone. Anthropic plans to open-source these optimizations.
Safety, Safeguards, and the Mythos Distinction
The key difference between Fable 5.1 and Mythos 5.1 lies in their safeguard layers. Fable 5.1 applies production-grade safety classifiers that redirect certain dual-use tasks to Opus-class models. Mythos 5.1 removes those restrictions for vetted users.
Cybersecurity Safeguards
Fable 5.1 can now be used to identify software vulnerabilities in source code, a capability that was previously restricted. Exploit development, penetration testing, and binary-based vulnerability scanning still redirect to Opus models. This change reduces safeguard interventions in Claude Code by roughly 60% per session. Biology safeguards fire 85% less often on benign elementary biology and medical questions compared to the safeguards that launched with Fable 5.
Safety Evaluations
Prior to release, Anthropic and external researchers subjected both models to extensive testing across chemical and biological risks, cyber capabilities, agentic safety, and alignment. Key findings include:
- Mythos 5.1’s chemical and biological capabilities remain below the next risk tier defined in Anthropic’s Responsible Scaling Policy
- No evidence of a critical-severity jailbreak was found for Fable 5.1’s safeguards during external testing
- Mythos 5.1 is better aligned across most metrics than its predecessor, with lower rates of reward hacking, motivated reasoning, and attempts to access resources outside its test environment
- The model is Anthropic’s most robust to date on external prompt injection benchmarks
Anti-Distillation Measures
Fable 5.1 introduces strengthened mechanisms against distillation attacks, a technique used to extract capabilities from advanced models. New API accounts can no longer manually edit Claude’s prior context in a multi-turn conversation while preserving the transcript of Claude’s thinking, closing off a publicly documented distillation technique. This change will apply to all users with future model releases.
EU AI Act Compliance
Fable 5.1 and Mythos 5.1 are the first Anthropic models subject to the EU AI Act’s Code of Practice on Transparency of AI-Generated Content. Outputs carry an invisible statistical text watermark with no impact on quality or content. Anthropic is rolling out a detection API in private preview for eligible organizations, including regulators, law enforcement, media, and educational institutions.
Enterprise Frontier Safeguards: Privacy Without Compromise
Alongside the new models, Anthropic announced Enterprise Frontier Safeguards (EFS), a system designed to solve a fundamental tension in enterprise AI adoption: the conflict between data privacy requirements and the need to monitor for misuse across sessions and accounts.
The Problem EFS Solves
When Anthropic introduced 30-day data retention with Fable 5, enterprises in regulated industries found it difficult to adopt the models despite understanding the safety value. Many operate under rules that tightly govern who may see certain information, from privileged legal material to non-public financial data to drug-safety reports. Adding another “trusted data vendor” required updating customer notifications, contracts, and internal audit processes.
EFS addresses this by storing activity data in cloud infrastructure the customer controls entirely. Data lives in the customer’s own Amazon S3, Azure Blob Storage, or Google Cloud Storage account, under their own encryption keys, access policies, and audit logging. Anthropic’s automated systems analyze a rolling window of traffic for signals of serious misuse, but flags go directly to the customer’s own team for review. No human review by Anthropic employees is required.
Enterprise Collaboration
Anthropic developed EFS with feedback from more than 100 customers spanning financial services, healthcare, manufacturing, telecom, law, retail, and the public sector. Partners include the Analysis and Resilience Center for Systemic Risk (ARC), whose members include CISOs from Goldman Sachs, Morgan Stanley, Citi, Bank of America, and Wells Fargo, as well as leaders at Comcast, KPMG, Mastercard, Salesforce, Stripe, and Visa.
EFS will be supported on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform, and Microsoft Foundry. It rolls out in phases starting fall 2026, with no additional charge from Anthropic. Until EFS is ready, eligible customers can use Fable 5.1 with zero data retention.
Availability, Platforms, and Breaking API Changes
Claude Fable 5.1 is available immediately on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. It is also accessible to Pro, Max, Team, and Enterprise users on Claude.ai. Claude Mythos 5.1 is currently limited to vetted U.S. organizations through the Cyber Verification Program and the Life Sciences Verification Program, with plans to expand access to international partners.
Breaking Changes for Developers
Teams migrating from Fable 5 should be aware of three breaking API changes:
- Forced tool use removed: Setting tool_choice to any or tool now returns a 400 error. Developers should use auto with strict tool use or structured outputs instead.
- Thinking blocks are model-bound: Fable 5.1 can read thinking blocks from earlier models, but earlier models cannot read Fable 5.1’s thinking blocks. Router and fallback setups that switch between models will lose reasoning context when downgrading.
- Editing earlier turns invalidates thinking: Injecting or deleting per-turn reminders, or rebuilding the system or tools array mid-conversation, now produces errors for accounts created on or after August 31, 2026. The fix is to use turn-scoped system messages and server-side context editing.
Anthropic also documents known regressions: parallel tool calling is more variable (agent loops may issue one call per turn instead of batching), the model narrates less and answers from memory more often at low effort, and it prefers whole-file rewrites over targeted edits.
Frequently Asked Questions
Is Claude Fable 5 still available?
Yes. Claude Fable 5 remains available on the Claude API and all supported platforms. Fable 5.1 is a successor, not a replacement, though Anthropic expects most users to migrate given the performance improvements and cost reductions.
Is Fable 5.1 included in Claude Max?
Yes. Claude Fable 5.1 is available to Pro, Max, Team, and Enterprise plan users on Claude.ai, as well as through the Claude API and cloud partners.
How does Fable 5.1 compare to Opus 5?
Fable 5.1 outperforms Opus 5 on most benchmarks while costing significantly less thanks to the cache read discount. Early adopters describe it as offering Opus-level intelligence at Fable-class pricing with faster response times. Anthropic’s next Opus release, Claude Opus 5.5, arrived on September 22 and reversed the hierarchy again: it matches Fable 5.1 on most work while costing about 40% less to run than Opus 5.
What is the difference between Fable 5.1 and Mythos 5.1?
They are the same underlying model. Fable 5.1 applies production safeguards that redirect certain cybersecurity and life-sciences tasks to Opus models. Mythos 5.1 removes those restrictions for vetted organizations in the Cyber Verification Program and Life Sciences Verification Program.
Can Fable 5.1 find software vulnerabilities?
Yes. Fable 5.1 is now permitted to identify software vulnerabilities in source code, which was previously restricted. However, it cannot develop exploits for those vulnerabilities. Exploit development, penetration testing, and binary-based vulnerability scanning still redirect to Opus models.
