Claude Fable 5.1 & Mythos 5.1: Benchmarks, Safety Upgrades, 45% Cost Savings

Introduction

Anthropic announces Claude Fable 5.1 and Claude Mythos 5.1, the same model with different safety guardrail levels. Fable 5.1 is generally available; Mythos 5.1 is offered through trusted access programs for cybersecurity work.

Customer Feedback

Pricing

Fable 5.1 reduces typical workload costs by ~25% vs Fable 5, and up to ~45% for highly agentic workloads, due to a 75% price cut for cache reads (now $0.25 per million tokens).

Data Retention

Enterprise Frontier Safeguards (EFS) provides zero-data-retention privacy while storing data on customer-controlled cloud infrastructure. EFS rolls out fall 2026; eligible customers can use zero-data-retention Fable 5.1 until then.

Safety Guardrails

Improved guardrails reduce false positives by 60% in cybersecurity. Fable 5.1 can now identify software vulnerabilities but not develop exploits.

New Performance Frontiers

Fable 5.1 sets new standards for coding, knowledge work, and long-running tasks. At Low/Medium effort levels, it matches or exceeds Fable 5 at lower cost. It avoids shortcuts and fixes root causes (e.g., Millennium's rare crash). Benchmark results (production guardrails enabled):

Benchmark                     Fable 5.1      Fable 5      Opus 5      GPT-5.6 Sol
Agentic Coding Terminal-Bench 4.0   55.8% (60.9% Mythos)   42.0%       52.3%       37.3%
Knowledge Work GDPval-AA v2       1853           1723         1824        1711
Computer Use OSWorld 2.0 (partial) 77.9%          72.9%        75.4%       —
Computer Use OSWorld 2.0 (strict)  41.7%          36.1%        39.6%       —
Multidisciplinary Reasoning HLE (no tools) 60.9%      57.8%        56.6%       —
Multidisciplinary Reasoning HLE (with tools) 65.0%     63.8%        63.6%       —
Business Workflow AutomationBench 31.4%          17.1%        26.9%       19.6%
Agentic Coding CursorBench 3.2.0  73.4%          70.5%        70.0%       67.2%

Guardrail interventions zeroed scores on some tasks; cybersecurity tasks handled by Opus 4.8.

Partner Feedback

"In internal benchmarks, Claude Fable 5.1 solves more coding problems than Fable 5 or Opus 5, and reaches state-of-the-art for trading intuition. It stays readable throughout long multi-step tasks." — Jane Street Capital, Craig Falls, Head of Quantitative Research

"We're migrating Devin's Opus 5 traffic to Fable 5.1 on launch day. It matches or slightly beats Fable 5 at lower cost per task, and new cache pricing makes Fable-tier models economically viable for workloads we kept on Opus." — Cognition, Walden Yan, Co-founder & CPO

"A piece of code crashed once every million runs; our team couldn't explain it in 4–5 years. Fable 5.1 decompiled a vendor library, matched it to a core dump, and traced the crash to a library defect." — Millennium, Damien, Senior Portfolio Manager

"Fable 5.1 built a complex prototype in ~3 days: researched all services/docs, proposed a novel extensible design, then ran unattended for hours with strong verification loops, delivering each stage with visual demos and proof of success." — MongoDB, Ron Sanzone, Senior Software Engineer

"Fable-tier intelligence, Opus-tier price, Sonnet-tier speed. ~2x faster than Opus 5, half the tokens — a clear upgrade for daily Opus users." — Every, Dan Shipper, CEO

"Fable 5.1 set a new high on our research suite, proposing a novel solution dimension that broke a previous bottleneck. Better at creative problem-solving and delivering 'aha' moments." — IMC, Marquis Wong, Chief AI Engineer

"Across all build failures tested, Fable 5.1 correctly identified root causes. Communication is more efficient: concise, easier-to-understand updates." — Red Hat, Josh Boyer, Distinguished Engineer

"Reviewed a clinical study already signed off by three other frontier models. Fable 5.1 found a missed flaw, insisted on more tests, and generated a new hypothesis that turned an abandoned dataset into a new research direction in one afternoon." — Rakuten, Felix Giovanni Virgo, Chief AI Engineer

"In a 30-day simulated enterprise eval with full Square tool access, Fable 5.1 achieved far higher token efficiency than Opus 5. We'll use it for complex multi-day whiteboard work so engineers stay fast." — Square (Block), Willem Avé, Global Head of Product

"For research, greenfield projects, or long-horizon tasks, I'd use Fable 5.1 as lead orchestrator. In a 38-hour unattended ML run, it diagnosed a label artifact, corrected it, launched six parallel overnight experiments, and returned results with next steps. Given an open prompt to find the highest-leverage unattended problem, it surfaced an ignored production alert, pulled logs, and proposed a fix." — Ramp, Dwight Temple, Senior ML Engineer

"Standout capability is writing: more readable, meaningful, and follows our style guides. In blind tests vs Fable 5, I preferred its output. In Canva Code it built a rhythm game with real music where beat-hits matched generated level difficulty — no other model achieved this." — Canva, Danny Wu, Head of AI

"Best presentation generator across all models tested, both for slide craft and thorough research answers. Also top on Citations eval for financial fact recall, and first to answer all parts of complex multi-part questions." — Hebbia, Aabhas Sharma, CTO

"A change spanning 8+ services across 3 codebases: Fable 5.1 mapped the entire workflow end-to-end, from incoming calls down to individual functions, DB tables, and rows — accurately throughout." — Plaid, Aditya Gupta, Senior Software Engineer

"Evaluators preferred Fable 5.1 answers ~2:1 over Fable 5 across everyday QA, high-intent research, and drafting/artifact work, with significantly better grounding. Fable 5.1 is now our default recommendation wherever we used Fable 5." — Glean, Nilesh Dalvi, Head of Engineering

"On our hardest problems, Fable 5.1 shows a significant edge. On an 18-month 'grand challenge' it made substantive progress, avoiding shallow 'stamp-collecting' exploration and making clear blank-slate connections I haven't seen elsewhere. Also optimized a compute kernel Fable 5 couldn't, gaining ~35% speedup." — iGent, Sean Ward, Co-founder & CEO

"On our toughest browser-agent benchmark, Fable 5.1 completed 82% of tasks (~10 min each) vs Opus 5 74% and Fable 5 57%, using fewer tokens. Stronger than Fable 5 on every dimension, and never overshot critical stop points across hundreds of tasks." — Browserbase, Miguel Gonzalez, Tech Lead

"On internal Finance bench, Fable 5.1 matched Fable 5 accuracy with 20% fewer tokens. Significant slide-gen improvement: better at explaining complex data in plain English and producing banker-grade visualizations." — Rogo, Alex Wang, Applied AI

"Fable 5.1 handles long unattended work better. I've run workflows for extended periods; it never drifts: self-logs progress, reprioritizes when context shifts, and resumes exactly from interruptions." — Shopify, Ben Lafferty, Senior Director of Engineering

"On RedlineBench (contract redlining), Fable 5.1 jumped from 47.9 to 57.0. Most gains from first-pass quality (score doubled) and counterparty acceptance. Edits are more concise, changing less of the document on average." — Crosby, Raymond Lin, Member of Technical Staff

"Leading model on our incident investigation eval using real production incidents to test Bits Investigation's root-cause analysis. Fable 5.1 showed stronger reasoning than Opus 5 and diagnosed the most complex incident we've tested." — Datadog, Daniel Shan, Staff Engineer

"Strongest model we've tested on CursorBench 3.2 (73.4% at max effort). Especially good at verifying its own work, enabling it to finish hard coding tasks end-to-end." — SpaceXAI, Sualeh Asif, Director of ML

"On FrontierFinance (real investor workflows), Fable 5.1 rose from 49.2% to 55.9%. Improvement stems from deeper retrieval of credible sources: on an earnings-call question it pulled the transcript and extracted exact management figures while others relied on secondary reports." — Samaya, Yuhao Zhang, Head of Research

Safety, Security, and Alignment

Cyber Risk

Mythos 5.1 demonstrates the strongest cyber capabilities of any released model, yet remains in the lower-risk tier of the Frontier Compliance Framework. Fable 5.1 guardrails were stress-tested by Anthropic, two external firms, and Gray Swan automated tests; no critical-severity jailbreaks found.

Agentic Safety

Mythos 5.1 refuses malicious agentic coding and computer-use requests at rates comparable to prior models; it is the most robust on external prompt injection benchmarks to date.

Alignment

Automated behavioral audits show Mythos 5.1 better aligned than Mythos 5: less likely to access out-of-scope resources on impossible tasks, less motivated reasoning, less likely to ignore explicit constraints. Training data review shows lower reward-hacking attempt and success rates. Limitations: audits have less visibility into very long contexts and multi-agent setups; impossible-task coverage is limited but improving.

Enterprise Frontier Safeguards (EFS)

EFS enables zero-data-retention privacy while detecting/responding to misuse. Data stays on customer cloud; human review defaults to customer. Developed with 100+ customers across finance, healthcare, manufacturing, telecom, legal, retail, public sector, and cloud partners AWS, Google Cloud, Azure. Supported on Claude Code, Claude Enterprise, Claude Platform, Amazon Bedrock, Google Agent Platform, Microsoft Foundry. Phased rollout begins fall 2026.

Precise Cybersecurity Guardrails

Fable 5.1 guardrails updated for precision: ~60% fewer interventions per session in Claude Code. Now allows vulnerability identification (defensive work) but routes dual-use tasks (pen testing, exploit generation, binary vulnerability scanning) to Opus.

Anti-Distillation Mechanisms

Fable 5.1 includes enhanced anti-distillation measures. New API accounts (created today onward) cannot manually edit Claude's prior context in multi-turn conversations while retaining reasoning records, closing a known distillation technique. Rollout is phased; existing accounts unaffected until future releases.

Claude Mythos 5.1 Trusted Access

Mythos 5.1 available via two programs: Cyber Verification Program (CVP) for defensive security work, and a second program for vetted individuals/organizations. Also powers Claude Security product for code vulnerability scanning and patch suggestions.

EU AI Act Compliance

Anthropic signed the EU AI Act Code of Practice July 2026. Requires watermarking for models released after Aug 2, 2026 — an invisible statistical method to detect Claude-generated text, no quality impact, no user info. Private preview detection API for eligible organizations (regulators, law enforcement, media, fact-checkers, researchers, educators, EU civil society, enterprises with compliance needs). Access will expand over time.

Cost and Availability

Fable 5.1 available today on all platforms (AWS, Google Cloud, Azure). API model ID: claude-fable-5-1. Cache reads now $0.25 per million tokens (75% reduction). Typical workloads ~25% cheaper vs Fable 5; highly agentic workloads up to ~45% cheaper. Other pricing unchanged: $10/million input tokens, $50/million output tokens. Mythos 5.1 initially for vetted US organizations; expanding to broader domestic/international partners.

Cache reads

All other tokens

Indexed cost comparison based on four weeks of August 2026 production usage at default effort levels. Typical workloads cover Claude Enterprise, Claude Code, and API. Highly agentic workloads are context- and tool-heavy with cache reads dominating cost.