Claude Fable 5 Launches: Anthropic's First Public Mythos-Class AI Model

Anthropic released Claude Fable 5 on June 9, 2026 — the first Mythos-class model available to the public — achieving 80.3% on SWE-Bench Pro and 10%+ gains over Opus 4.8 across key benchmarks.

#Claude#Anthropic#Fable 5#Mythos#LLM

Claude Fable 5 Launches: Anthropic's First Public Mythos-Class AI Model

AI Summary

Anthropic released Claude Fable 5 on June 9, 2026 — the first Mythos-class model available to the public — achieving 80.3% on SWE-Bench Pro and 10%+ gains over Opus 4.8 across key benchmarks.

Introduction

On June 9, 2026, Anthropic released Claude Fable 5, marking a historic first: a Mythos-class model made available for general use. Previously, the Mythos tier existed only as a restricted capability deployed under Project Glasswing for vetted cybersecurity and government partners. Fable 5 brings that level of intelligence to the Claude API and to claude.ai subscribers, accompanied by layered safety classifiers designed to contain its most powerful capabilities. The companion model, Claude Mythos 5, shares the same underlying weights but operates with those classifiers removed — it remains restricted to a small group of Project Glasswing partners and trusted biomedical researchers.

Feature Overview

SWE-Bench Pro: 80.3% Coding Performance

Fable 5 scored 80.3% on SWE-Bench Pro, the industry benchmark for real-world software engineering tasks, compared to 69.2% for Claude Opus 4.8, 58.6% for GPT-5.5, and 54.2% for Gemini 3.1 Pro. Stripe reported that Fable 5 compressed a 50-million-line codebase migration that would have taken two months of manual engineering effort into a single day. For development teams running large-scale autonomous coding workflows, this benchmark gap translates directly into measurable productivity gains.

Knowledge Work: GDPval-AA 1932

On GDPval-AA, a compound benchmark measuring multi-step research synthesis and analytical reasoning, Fable 5 scored 1932 compared to 1890 for Opus 4.8, 1769 for GPT-5.5, and 1314 for Gemini 3.1 Pro. The model demonstrates consistent improvement on tasks requiring self-validation across many sequential reasoning steps — the type of work involved in legal document review, scientific literature synthesis, and strategic analysis.

Long-Context and Vision

Fable 5 maintains focus coherently across millions of tokens, a capability Anthropic demonstrated through a game-playing benchmark where memory integration tripled performance in Slay the Spire compared to models without long-context recall. On Blueprint-Bench 2, which measures spatial reasoning and visual code reconstruction from screenshots, Fable 5 scored 38.6% — ahead of GPT-5.5 (36.2%) and Gemini 3.1 Pro (26.5%), and dramatically higher than Opus 4.8 (14.5%). It can extract precise numerical data from scientific figures and reconstruct functional code from interface screenshots.

Three-Domain Safety Architecture

The central design decision in Fable 5 is its tripartite safety classifier system. Three domains trigger automatic rerouting to Claude Opus 4.8:

Cybersecurity — Exploitation guidance, offensive cyber operations, and attack tooling
Biology/Chemistry — Dual-use biological and chemical synthesis requests, with plans to narrow the trigger threshold as calibration improves
Distillation — Attempts to extract the underlying model weights through systematic prompting

This fallback architecture means that the ExploitBench score of 78.0% and the strong biology benchmark results belong to Claude Mythos 5, not Fable 5. The version available through the public API performs closer to Opus 4.8 on those safeguarded domains. Anthropic conducted over 1,000 hours of external red-team testing and found no universal jailbreak, though the UK AI Safety Institute made partial progress toward constructing one.

Usability Analysis

Fable 5 is immediately accessible through the Claude API under the model identifier claude-fable-5. It is available on claude.ai, AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Subscription plan users received access in the June 9–22 window without requiring additional credits; from June 23 onward, subscription usage draws from a separate credit pool at metered API rates.

For developers evaluating whether to migrate from Opus 4.8, the decision comes down to task profile. Fable 5 justifies its 2x price premium for autonomous multi-day coding tasks across large codebases, complex multi-stage research synthesis requiring self-validation, and visual tasks involving diagram interpretation or code reconstruction. For high-volume classification, summarization, or latency-sensitive applications, Opus 4.8 remains the more cost-effective choice.

Prompt caching reduces the effective input cost by 90%, making Fable 5 more economically viable for workflows that reuse large context windows — a common pattern in agentic coding and document analysis.

Pros and Cons

Advantages:

State-of-the-art SWE-Bench Pro score of 80.3%, a significant lead over all competitors
First Mythos-class model accessible through a public API with broad platform support
Exceptional vision performance, including accurate extraction from scientific figures and UI screenshots
90% prompt caching discount mitigates the high base price for context-heavy workflows
Demonstrated real-world impact: Stripe's codebase migration from two months to one day

Limitations:

Priced at $10/$50 per million tokens — double the cost of Claude Opus 4.8 ($5/$25)
The most capable cybersecurity and biology scores belong to Mythos 5, not the publicly accessible Fable 5
Biology/chemistry safeguards described as broadly calibrated, which may trigger on legitimate research queries
Mythos 5 remains inaccessible to the general developer community
Full subscription inclusion reverts to credit-based access starting June 23

Outlook

Fable 5 represents a deliberate experiment in capability release: deploying maximum intelligence with surgical safety constraints rather than holding the model back entirely. Anthropic's approach of maintaining an identical Mythos 5 backend for restricted partners suggests the company views the classifier layer as modular — adjustable as alignment research matures rather than permanently baked into the architecture.

The immediate market impact is substantial. A gap of more than 20 points over GPT-5.5 on SWE-Bench Pro will pressure OpenAI and Google to accelerate their own flagship releases. The Stripe use case — months of migration work compressed to a day — provides the kind of concrete ROI narrative that enterprise sales teams will amplify broadly.

Longer term, the trusted-access expansion planned for vetted cybersecurity organizations and biomedical researchers will gradually bring Mythos 5's full capability into the hands of domain experts while maintaining governance guardrails that Anthropic can audit and adjust.

Conclusion

Claude Fable 5 is the most capable publicly accessible model as of June 2026, establishing new benchmarks in software engineering, knowledge work, and visual reasoning. Its three-domain safety architecture is an honest trade-off: general users gain access to Mythos-class intelligence, while the most sensitive capabilities remain gated. For development teams working on large-scale autonomous coding, complex research synthesis, or document understanding from visual sources, Fable 5 is a meaningful step-change. The $10/$50 pricing is real, but the demonstrated throughput gains make the economics defensible for the right workloads.

Editor's Verdict

Claude Fable 5 marks a new tier in publicly accessible AI. Anthropic's decision to open the Mythos class — even with safety classifiers in place — is the most significant model release of 2026 so far. Teams already invested in the Claude ecosystem should evaluate Fable 5 seriously for autonomous coding pipelines; everyone else should benchmark it against their specific workloads before committing to the premium price.

Pros

State-of-the-art coding performance: 80.3% SWE-Bench Pro, more than 20 points ahead of GPT-5.5
First Mythos-class model with public API access across Claude.ai, AWS, GCP, and Microsoft Foundry
Leading vision and spatial reasoning on Blueprint-Bench 2, enabling UI screenshot-to-code and scientific figure extraction
90% prompt caching discount makes long-context agentic workflows significantly more cost-efficient
Demonstrated enterprise ROI: Stripe compressed a two-month migration into one day

Cons

Priced at $10/$50 per million tokens — double the cost of Claude Opus 4.8, making it unsuitable for high-volume routine workloads
The top cybersecurity and biology benchmark scores belong to the restricted Mythos 5, not the publicly available Fable 5
Biology/chemistry safety classifiers are broadly calibrated and may generate false positives for legitimate research queries
Subscription-based free access ends June 22; metered credit usage required from June 23 onward

References

Claude Fable 5 and Mythos 5 — Anthropic Official Announcement Claude Fable 5 & Mythos 5: The Frontier, Split in Two — Digital Applied Anthropic's Claude Fable 5 Launch — IT Pro Claude Mythos Goes Public in Fable 5 — 9to5Google

Comments0

Key Features

1. SWE-Bench Pro score of 80.3% — the highest recorded for a publicly available model, 11 points ahead of Claude Opus 4.8 and 22 points ahead of GPT-5.5. 2. Three-domain safety classifier system automatically routes cybersecurity, biology/chemistry, and distillation queries to Claude Opus 4.8, enabling broad API access without unrestricted Mythos-level capability. 3. Blueprint-Bench 2 score of 38.6% demonstrates leading spatial and visual reasoning, enabling precise extraction from scientific figures and UI screenshot-to-code reconstruction. 4. Long-context coherence across millions of tokens with memory integration that tripled game-playing performance in controlled benchmarks. 5. Prompt caching at 90% input discount reduces the effective cost for workflows that reuse large context windows, partially offsetting the 2x price premium over Opus 4.8.

Key Insights

Claude Fable 5 scores 80.3% on SWE-Bench Pro, more than 20 points ahead of GPT-5.5 and 11 points ahead of Claude Opus 4.8 — the largest competitive gap in coding benchmarks since GPT-4 launched.
The tripartite safety classifier (cybersecurity, biology/chemistry, distillation) is the architectural innovation that made public release possible; the underlying Mythos 5 model scores 78.0% on ExploitBench with classifiers removed.
Stripe's real-world result — a 50-million-line codebase migration in one day versus two months manually — provides a concrete productivity multiplier that resonates with enterprise decision-makers.
Fable 5's Blueprint-Bench 2 score of 38.6% for spatial reasoning significantly exceeds GPT-5.5 (36.2%), Gemini 3.1 Pro (26.5%), and Opus 4.8 (14.5%), opening new use cases in diagram interpretation and visual code reconstruction.
The June 9–22 free access window on subscription plans, followed by credit-based metered billing from June 23, reflects Anthropic's strategy of driving rapid adoption before transitioning to premium monetization.
At $10/$50 per million tokens, Fable 5 costs twice as much as Opus 4.8; prompt caching at 90% discount is the primary lever for making agentic workflows economically viable.
Claude Mythos 5 — the same model without safety classifiers — remains restricted to Project Glasswing partners and a small trusted-access program for biomedical researchers, with expansion planned but no timeline given.
The UK AI Safety Institute's partial progress toward constructing a universal jailbreak during external red-teaming suggests the classifier architecture is robust but not impenetrable — an important caveat for organizations in sensitive industries.