Claude Fable 5

Anthropic
Model overview

claude-fable-5

Anthropic's most capable widely available model for demanding reasoning and long-horizon agentic work.

general LLMreasoningcodingvision-languageagentcomputer uselong contextchatcode generationsoftware engineering agentdata analysisdocument understandingenterprise access
Model risk

Current observations

4current adverse observations
Negative / riskData / governanceData retentionOfficial · Active

Safety safeguards require 30-day retention and do not support zero-data retention

Anthropic's Fable 5 safeguard documentation requires 30-day safety retention and states that zero-data retention is unavailable, directly affecting regulated data and sensitive-code deployments.

Introducing Claude Fable 5 and Claude Mythos 5High confidenceObserved 2026-07-01Source checked 2026-07-15Current state verified 2026-07-15
Action

Confirm data classification with legal and security teams; choose another model when ZDR is mandatory.

Verify source
Negative / riskModel behaviorSafety robustnessResearch · Active

A red-team study still observed jailbreak success under automated attacks

A preprint reported a 6.1% worst-case attack success rate for Fable 5 under its test setup. The figure applies only to that methodology and sample, not a general real-world risk rate.

A Red-Team Study of Anthropic Fable 5 & Opus 4.8 ModelsMedium confidenceObserved 2026-06-20Source checked 2026-07-15
Action

High-risk deployments still require independent red teaming and output isolation.

Verify source
Negative / riskAPI / serviceGuardrails & refusalsOfficial · Active

Safety classifiers can return refusals with HTTP 200, requiring explicit fallback handling

Official integration guidance requires applications to detect stop_reason=refusal and retry through server-side, SDK, or manual fallback paths.

Introducing Claude Fable 5 and Claude Mythos 5High confidenceObserved 2026-06-09Source checked 2026-07-15Current state verified 2026-07-15
Action

Ignoring the refusal branch can misclassify a successful HTTP response as completed work.

Verify source
Negative / riskProduct experienceGuardrails & refusalsNews · Monitoring

Security researchers reported guardrails triggering on routine code review

TechCrunch summarized reports from security researchers that some routine code-review and security-reading requests triggered Fable safeguards or fallback behavior.

Cybersecurity researchers aren't happy about the guardrails on Anthropic's FableMedium confidenceObserved 2026-06-10Source checked 2026-07-15Review after 2026-07-31
Action

This is feedback from a specific professional cohort, not all workloads; security teams should run workflow-specific trials.

Verify source
Historical records 1
HistoricalAPI / serviceAvailabilityOfficial · Historical

Access was suspended under export controls and restored globally on July 1

Access was suspended on June 12 and restored globally on July 1. The interruption ended and is retained as a supply-continuity record.

Redeploying Claude Fable 5High confidenceObserved 2026-07-01Source checked 2026-07-15Moved to history 2026-07-01
Action

No active outage response is required; retain provider-switching and model-fallback plans.

Verify source
Representative version

claude-fable-5

CompanyAnthropic
Release2026-06-09
Release typeModel release
ParametersNot disclosed
ArchitectureNot disclosed
Context1,000,000
Input modalitiestext / image
Output modalitiestext
Tool use / function callingConfirmed: Tool use, Function calling
Structured output / reasoningConfirmed: Structured output, Reasoning
Vision / audio / videoConfirmed: Vision; Not shown: Audio, Video
Open weights / LicenseNo open weights / license not applicable
anthropic-claude-fable-5-current · License not applicable
Vendor APIAvailable
Introducing Claude Fable 5 and Claude Mythos 5anthropic-claude-fable-5-current · 2026-06-09
OpenRouterOpenRouter listed
PricingOfficial · 2026-06-09: $10 / $50
OpenRouter · 2026-07-15: $10 / $50
input / output, per 1M tokens
BenchmarkNo reliable public source
Public signalsNo reliable public source
Use casesLong-horizon agents / Complex coding / Enterprise knowledge work / Vision analysis
Claim evidenceRelease / model specs / weightsIntroducing Claude Fable 5 and Claude Mythos 5claude-fable-5 · 2026-06-09
OpenRouter

Pricing / context / availability

OpenRouterOpenRouter listed
OpenRouter listinganthropic/claude-fable-5
Mapped versionclaude-fable-5
anthropic-claude-fable-5-current
PricingOfficial · 2026-06-09: $10 / $50
OpenRouter · 2026-07-15: $10 / $50
input / output, per 1M tokens
Context1000000
Speed/LatencyNo reliable public source
Snapshot time2026-07-15T20:11:41.551Z
SourceView source
Evidence gaps

Public display boundary

Parameter scale not disclosedNo reliable benchmark sourceNo public model-card activity countersNo linkable real reviewSpeed/latency not disclosed