The Coe Lab
← Back to Blog

Claude Fable 5: The Mythos-Class Model Goes Public

June 14, 20268 min read

Anthropic's most capable public model yet. Fable 5 brings Mythos-level reasoning to everyone, with benchmarks that beat the competition and autonomous capabilities that change what's possible.

Claude Fable 5: The Mythos-Class Model Goes Public

On June 9, 2026, Anthropic made a move that reshapes the AI landscape. They released Claude Fable 5—their most capable generally available model ever. This isn't just another incremental update. It's the public debut of what Anthropic calls their "Mythos-class" capability tier, a level of reasoning and autonomy that until now was restricted to vetted partners working on cybersecurity defense.

What Makes Fable 5 Different

The story begins in April 2026, when Anthropic released Claude Mythos Preview through Project Glasswing. The model was unexpectedly powerful at finding and exploiting software vulnerabilities—so powerful that Anthropic restricted access to cyber defenders and infrastructure providers, working with the US government to ensure responsible deployment.

Fable 5 is the same underlying model, wrapped in new safety systems that make broad release possible. The name itself is meaningful: "fable" comes from the Latin fabula ("that which is told"), which shares roots with the Greek mythos. The capability is the same; only the safety filters differ.

Benchmark Performance That Matters

The numbers are striking. Independent testing firm Artificial Analysis ranked Fable 5 number one on their Intelligence Index with a score of 64.9—about 5 points ahead of GPT 5.5, the closest competitor. But raw benchmarks only tell part of the story.

  • SWE-Bench Pro: 80.3%—real-world bug fixes in open-source projects
  • SWE-Bench Verified: 95.5%—human-verified coding accuracy
  • Terminal-Bench 2.1: 88%—command-line task completion
  • OSWorld-Verified: 85%—operating computers autonomously
  • GDPval-AA: 1932—economically valuable knowledge work, rated like chess Elo

Fable 5 scored more than 10% higher than Claude Opus 4.8 on some benchmarks, with the biggest gains in software engineering and long-running agent tasks—the exact areas where Anthropic was already strongest.

Real-World Capabilities

The demonstrations are where Fable 5's capabilities become tangible:

  • Stripe reported Fable 5 compressed months of engineering work into days—performing a codebase-wide migration on a 50 million line Ruby codebase in a single day, work that would have taken a team over two months by hand.
  • Vision capabilities that can pull precise numbers from scientific figures and rebuild a web app's source code from screenshots alone.
  • Memory that scales: On the Slay the Spire benchmark, giving Fable 5 file-based memory improved its performance three times more than it did for Opus 4.8.
  • Autonomous game completion: Earlier Claude models couldn't finish Pokémon FireRed even with helper tools. Fable 5 beat the game using nothing but raw screenshots.

The model can also autonomously play Factorio, design 3D-printable CAD models in browser-based editors (and build those editors from scratch), and create fluid simulations synced to music it composed through code.

The Safety System Architecture

This is where Fable 5 gets interesting for anyone building production systems. Anthropic has built separate AI classifiers that watch for misuse in three specific areas:

  • Cybersecurity—Blocking vulnerability exploitation and multi-step "agentic" hacking attempts
  • Biology and Chemistry—Currently a broad net that routes most bio/chem queries to Opus 4.8, designed to narrow over time
  • Distillation—Preventing large-scale copying of Claude's capabilities to train competing models

When a request trips one of these classifiers, Fable 5 doesn't refuse—it falls back to Claude Opus 4.8. The user gets a real answer from a strong model, just not the Mythos-level one. Anthropic's early data shows over 95% of Fable sessions involve no fallback at all.

The company also ran an external bug bounty that found no universal jailbreak in over 1,000 hours of testing, while acknowledging that fully preventing jailbreaks is likely impossible—the goal is making them slow and costly enough to catch.

Pricing and Technical Details

For developers building on Fable 5:

  • Model ID: claude-fable-5
  • Context window: 1 million tokens
  • Maximum output: 128K tokens per request
  • API pricing: 0 per million input tokens, 0 per million output tokens
  • Availability: Anthropic API, Amazon Bedrock (US East, Europe Stockholm), Google Vertex AI, Microsoft Foundry

A key technical difference: thinking is always on. You cannot disable it, and the raw chain of thought is never exposed. Instead, you control thinking depth through an effort parameter. Temperature, top_p, top_k, and budget_tokens parameters are not accepted.

The Data Retention Tradeoff

This is the catch you need to know about. Fable 5 is designated a "Covered Model," which means Anthropic requires 30-day retention of inputs and outputs for all traffic. Zero data retention is not available.

Anthropic says it won't use this data for training and deletes it after 30 days in most cases. The purpose is catching complex attacks that span multiple requests. On AWS Bedrock, you must explicitly opt into data sharing before using the model—data that leaves the AWS security boundary.

For companies with strict data residency or privacy requirements, this is a real consideration.

The June 22 Deadline

For subscription users (Pro, Max, Team, Enterprise), there's an important window: Fable 5 is included at no extra cost through June 22, 2026. After that date, using it requires usage credits until Anthropic has enough capacity to restore it as a standard plan feature.

If you want to test the model on your existing subscription, do it now.

What About Mythos 5?

Mythos 5 is the unrestricted version—same model, lifted cyber safeguards. It's available only through Project Glasswing approval, in consultation with the US government. This is where the most striking science results come from:

  • Drug design: Anthropic's protein design experts reported 10x speedup on parts of the drug design process. Nine of 14 protein targets yielded strong drug candidates now under investigation.
  • Genomics: Over a week of mostly autonomous work, Mythos 5 trained a model that outperformed a recent Science journal model despite being 100x smaller.
  • Novel biology: In blind comparisons, Anthropic scientists preferred Mythos 5's hypotheses about 80% of the time. One hypothesis about E. coli proteins was independently confirmed by another lab.

The Bottom Line

Claude Fable 5 is the most capable model available to the public right now. The lead on coding and autonomous work is real and independently verified. But the tradeoffs are equally clear: premium pricing, mandatory data retention, a safety system that quietly routes some requests to a weaker model, and subscription access that pauses after June 22.

For teams doing serious software engineering, complex document analysis, or long-running automated work, Fable 5 is worth testing immediately. For everyday tasks, models like Sonnet 4.6 still cover most needs at a fraction of the cost.

The real story isn't just benchmarks—it's that Anthropic has figured out how to safely release capability levels that were previously restricted to government partners. The safety-first approach that made them famous is now enabling faster, more confident deployment of their most powerful models. That's a shift worth watching.

References