update · TowCue Editorial Team

Anthropic launches Claude Fable 5.1 and Mythos 5.1: performance gains, cache read price cut, enterprise safeguards

Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, 2026 — same model, split safeguards. Fable 5.1 is GA with 25–45% cost reduction via 75% cheaper cache reads, Mythos 5.1 for vetted cyber/life-science work. Enterprise Frontier Safeguards (EFS) rolling out this fall for zero-data-retention privacy with automated misuse detection.

Original editorial contentSources verifiedLast reviewed: 2026-09-02

Quick answer

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026 — the same underlying model with split safeguard tiers.

Fable 5.1 (GA): Available now on all platforms (Claude API, AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure). ~25% cheaper for typical workloads, up to ~45% cheaper for highly agentic work — driven by a 75% cache read price cut (now $0.25/M tokens). Base pricing unchanged: $10/M input, $50/M output.

Mythos 5.1 (Restricted): Same model, more permissive safeguards for vetted cyberdefenders and life-science researchers. Access via Cyber Verification Program (CVP) and Life Sciences Verification Program (LSVP, US-government partnership). Currently US-only; expansion planned.

Enterprise Frontier Safeguards (EFS): Rolling out this fall — zero-data-retention privacy (customer-controlled cloud storage) + automated misuse detection. Developed with 100+ enterprise customers. Supported on Claude Code, Enterprise, Platform, Bedrock, Vertex AI, Foundry, Google Agent Platform.

EU AI Act compliance: Watermarking on post-August-2 outputs; detection API in private preview for eligible orgs.

Model ID: claude-fable-5-1 on Claude API. Available on AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure.


What changed

Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026 — the same underlying model with different safeguard profiles:

TierNameSafeguardsAvailability
GAFable 5.1Standard production safeguardsAll platforms, all customers
RestrictedMythos 5.1Permissive cyber/life-science safeguardsVetted orgs only (CVP, LSVP)

Key insight: They are the same model weights; the difference is purely in safeguard configuration.


Performance gains

Fable 5.1 sets a new bar for coding, knowledge work, and long-horizon problem solving. At Medium effort, it often matches or beats Fable 5 at High effort — at much lower cost.

BenchmarkFable 5.1Fable 5Opus 5GPT-5.6 Sol
Agentic scientific research (Terminal-Bench-Science 0.1)52.6%24.7%29.0%22.4%
Agentic coding (Terminal-Bench 4.0)55.8%42.0%52.3%37.3%
Knowledge work (GDPval-AA v2)1853172318241711
Computer use (OSWorld 2.0 accuracy)77.9%72.9%75.4%
Multidisciplinary reasoning (Humanity's Last Exam, no tools)60.9%57.8%56.6%
Agentic coding (CursorBench 3.2.0)73.4%70.5%70.0%67.2%

Key qualitative improvements (from early-access partners):

  • Readable over long tasks — "remains readable over long, multi-step tasks" (Jane Street)
  • Root-cause analysis — found a 1-in-a-million crash in a vendor library (Millennium)
  • Unattended multi-day work — 38-hour ML run diagnosed label artifact, launched 6 parallel experiments (Ramp)
  • 2× faster than Opus 5, half the tokens (Every)
  • Creative problem solving — novel solution on a grand-challenge problem (iGent)
  • Contract redlining — 47.9 → 57.0 on RedlineBench, first-turn quality doubled (Crosby)

Pricing: the cache read revolution

MetricFable 5Fable 5.1Change
Input tokens$10/M$10/M
Output tokens$50/M$50/M
Cache reads$1.00/M$0.25/M–75%

Why it matters: Cache reads dominate cost in context-heavy, tool-heavy, multi-turn agentic work (long context, repeated reads, computer use, code execution).

Workload typeCost reduction vs Fable 5
Typical (Enterprise, Code, API mix)~25%
Highly agentic (context-heavy, tool-heavy, computer use)~45%

Example: A 38-hour unattended ML run with 6 parallel overnight experiments — the kind of workload where cache reads dominate — sees ~45% cost reduction.

Base pricing unchanged: $10/M input, $50/M output. Cache reads now $0.25/M (was $1.00/M).


Mythos 5.1: the restricted twin

Same model, permissive safeguards for two vetted communities:

ProgramFocusStatus
Cyber Verification Program (CVP)Defensive security: vulnerability discovery, defensive codingExpanding to Mythos-class models
Life Sciences Verification Program (LSVP)Professional R&D in biology/medicine (US-gov partnership)First participants enrolled; expanding

Mythos 5.1 capabilities (per System Card):

  • Chemical/biological: CB-1 level (can help synthesize known weapons with basic technical background) — below CB-2 (replacing rare expert talent). Same safeguards as Mythos 5.
  • Cyber: Strongest cyber capabilities of any released model; still lower risk tier in Frontier Compliance Framework. No critical-severity jailbreak found.
  • Agentic safety: Most robust to prompt injection to date (external benchmark).
  • Alignment: Better than Mythos 5 — less resource-exfiltration attempts, less motivated reasoning, less constraint-ignoring. Still: can sometimes bypass approvals/auto-mode classifiers; limited visibility into very long-context & multi-agent settings.

Claude Security (vulnerability scanner) now runs on Mythos 5.1.


Enterprise Frontier Safeguards (EFS): zero data retention + automated detection

FeatureDescription
Data locationCustomer-controlled cloud (AWS, GCP, Azure) — not Anthropic
Human reviewBy customer, not Anthropic (default)
PrivacyEquivalent to zero data retention (ZDR)
Misuse detectionAutomated, state-of-the-art
RolloutPhased, starting fall 2026
BridgeEligible customers get ZDR on Fable 5 / 5.1 until EFS ready

Supported on: Claude Code, Claude Enterprise, Claude Platform, AWS Bedrock, Claude on AWS, Google Vertex AI, Google Agent Platform, Microsoft Foundry.

Developed with: 100+ enterprise customers (finance, healthcare, manufacturing, telecom, law, retail, public sector) + cloud partners (AWS, GCP, Azure).


Safeguard improvements (Fable 5.1)

AreaChange
Biology false positives85% fewer interventions on benign elementary biology/medical queries
Cyber false positives~60% fewer safeguard interventions per session in Claude Code
Vulnerability discoveryNow allowed (defensive work) — exploit gen, pen testing still redirected to Opus
Anti-distillationNew API accounts: cannot manually edit prior context while preserving thinking transcript (blocks common distillation technique). Rolling out gradually.

EU AI Act compliance (effective Aug 2, 2026)

RequirementImplementation
WatermarkingInvisible text watermark on all post-Aug-2 model outputs
Detection APIPrivate preview for regulators, law enforcement, media, fact-checkers, researchers, EU civil society, compliance-obligated enterprises
No quality impactWatermark invisible, no user/org/conversation metadata

Benchmark deep dive (selected)

BenchmarkFable 5.1Fable 5Notes
Terminal-Bench-Science 0.1 (agentic science)52.6%24.7%2.1× gain
Terminal-Bench 4.0 (agentic coding)55.8% (Fable) / 60.9% (Mythos)42.0%Mythos gap = safeguard interventions
CursorBench 3.2.0 (agentic coding)73.4%70.5%Best Fable-class score
OSWorld 2.0 (computer use, accuracy)77.9%72.9%
Humanity's Last Exam (no tools)60.9%57.8%
AutomationBench (business workflows)31.4%17.1%
RedlineBench (contract redlining)57.047.9First-turn quality doubled

Note: Fable 5.1 evaluated with production safeguards on. Tasks where safeguards intervened score zero on OSWorld 2.0 / AutomationBench; cyber tasks completed by Opus 4.8, biology by Opus 5.


Availability matrix

PlatformFable 5.1Mythos 5.1
Claude APIclaude-fable-5-1CVP/LSVP only
Claude.ai / Claude Code / Cowork✅ Default: Medium effort (Code: High)CVP/LSVP only
AWS BedrockCVP/LSVP only
Google Cloud Vertex AICVP/LSVP only
Microsoft Azure / FoundryCVP/LSVP only
Claude Security✅ Now powered by Mythos 5.1

Effort defaults: Fable 5.1 defaults to High in Claude Code, Medium in Cowork and Claude.ai.


What we don't know / unconfirmed

  • Exact Mythos 5.1 global rollout timeline (US-only now; "coordinating with US government to expand")
  • EFS exact GA date ("phased, starting this fall")
  • Detection API public GA timeline ("expand access over time")
  • Anti-distillation impact on existing custom integrations ("small number affected")
  • Long-context / multi-agent alignment coverage (acknowledged gap in System Card)
  • Whether cache read discount applies to Bedrock/Vertex/Foundry (API pricing mentioned; cloud partner pricing not explicitly confirmed)
  • Opus 5.1 / Sonnet 5.1 timeline ("bringing many improvements to rest of model family")

TowCue take

This release is three stories in one:

  1. Economics shift — The 75% cache read cut is the real story. It makes Fable-class models finally economical for agentic workloads (Cognition moving Opus traffic to Fable 5.1 on launch day). For teams running context-heavy agents (computer use, code execution, long-context RAG), this is a step-function cost reduction.

  2. Safeguard-as-product-differentiation — Same weights, two SKUs via safeguard config. Mythos 5.1 for vetted pros; Fable 5.1 for everyone else. This is a regulatory/compliance product strategy — expect more of this as AI regulation matures.

  3. Enterprise privacy architecture — EFS = ZDR + automated detection, data in your cloud. This is the privacy architecture enterprises have been waiting for — no Anthropic human review, data never leaves your VPC. If you're in regulated industry (finance, healthcare, gov), this is the feature to watch.

For developers: Fable 5.1 is now the default choice for agentic coding — better than Fable 5, cheaper than Opus, readable over long runs. Switch Cursor/Code traffic today.

For enterprises: If you need ZDR + misuse detection, get on the EFS waitlist now. Bridge ZDR on Fable 5/5.1 is available immediately for eligible customers.

For regulated industries: Mythos 5.1 via CVP/LSVP is now the path for defensive cyber / life-science R&D. Start the vetting process if you're in those spaces.

For compliance: Watermarking + detection API = EU AI Act compliance checklist item. If you're an enterprise with AI-generated content obligations, register for detection API access.

Bottom line: Fable 5.1 makes agentic workflows economically viable at scale. The cost curve just bent. If you've been holding back on multi-turn, tool-heavy agents due to cost — the barrier just dropped ~25–45%.

Verify current pricing, platform availability, and program access at Anthropic's Fable 5.1 page.

Research sources

Related decision guides