AI & Technology

Introducing Claude Fable 5 and Claude Mythos 5: Capabilities, API Changes, and Availability

By Khushal Charaniya · June 12, 2026 · 10 min read · Technology

What Are Claude Fable 5 and Claude Mythos 5?

Anthropic's fifth-generation model family consists of two variants that share the same underlying neural network but differ in how their safety guardrails are applied.

The key distinction: Claude Fable 5 includes safety classifiers that block high-risk queries, falling back to Claude Opus 4.8 when triggered. Claude Mythos 5, deployed through Project Glasswing, removes those restrictions in areas like cybersecurity — making it powerful enough to both identify and defend against sophisticated software vulnerabilities.

Benchmark Performance: How Fable 5 Compares

Numbers tell the clearest story here. Fable 5 doesn't just edge out competitors — it wins by a margin wide enough to matter in real work.

On Cognition's FrontierCode evaluation — which tests whether AI can pass demanding coding tasks while meeting the standards of production codebases — Fable 5 scores highest among all frontier models, even at medium effort settings. That 29.3% versus Opus 4.8's 13.4% is not a narrow gap.

Key Capabilities Breakdown

Software Engineering

This is where Fable 5 makes the most striking impression. Stripe ran a real-world test: migrating a 50-million-line Ruby codebase. With Fable 5, the migration was done in a day. Without it, the same task would take a whole team more than two months.

The model also demonstrated superior token efficiency — doing more with less output, which matters enormously in production agentic workflows. Cursor's CEO called Fable 5 "state of the art" and said it "opened up a class of long-horizon problems that were out of reach for earlier models."

Vision Capabilities

Fable 5 handles vision tasks that earlier Claude models simply couldn't manage without heavy scaffolding. One notable example from Anthropic: earlier Claude models required complex helper tools to play Pokémon FireRed at all. Fable 5 completed the game using only raw game screenshots — no maps, no navigation aids, nothing beyond what a human player would see on screen.

The model can also extract precise numbers from dense scientific figures and rebuild a web app's source code from screenshots alone.

Knowledge Work and Finance

On Hebbia's Finance Benchmark — which tests senior-level analytical reasoning — Fable 5 achieved the highest score of any model tested. IMC, a trading firm, reported that Fable 5 "aced their trading-analysis evaluations nearly across the board," including factual lookup, root-cause analysis, and expected-value calculations.

Memory and Long-Context Performance

Fable 5 operates with a 1 million token context window by default and improves its outputs by taking and referencing its own notes during long tasks. In a test using the card game Slay the Spire, giving Fable 5 persistent file-based memory improved performance three times more than the equivalent setup did for Opus 4.8.

Life Sciences and Drug Design (Mythos 5)

Anthropic's internal protein design team used Mythos 5 to accelerate aspects of drug design by roughly ten times. In one experiment, Mythos 5 worked autonomously — choosing binding sites, selecting tools, and recovering from failures — on 14 protein targets. Nine of the 14 yielded strong candidates that the team is currently investigating.

Safety Classifiers: What Developers Need to Know

This is the most important change for developers integrating Fable 5 into production systems. The model includes safety classifiers that can refuse requests, and when they fire, the API behavior is different from what you may be used to.

• 1
  Refusals return HTTP 200, not an error

  When Fable 5 declines a request, the Messages API returns stop_reason: "refusal" as a successful response. It also reports which classifier triggered the block. You need to check for this in your response handling.

• 2
  Server-side fallback is available

  Pass the fallbacks parameter in your API call and the server will automatically retry refused requests on Claude Opus 4.8 or another available model. This is currently in beta.

• 3
  Client-side fallback via SDK middleware

  SDK middleware is available for TypeScript, Python, Go, Java, and C# to handle retries on the client side, compatible with any platform.

• 4
  Classifiers fire in less than 5% of sessions on average

  Anthropic tuned these conservatively, which means some harmless requests will occasionally be blocked. They are working to reduce false positives as more capable models arrive. An external bug bounty produced no universal jailbreaks across 1,000+ hours of testing.

High-risk areas that trigger the classifiers include cybersecurity, biology, chemistry, and distillation. When triggered in these areas, the model falls back to Claude Opus 4.8 rather than refusing entirely.

Availability and Pricing

For subscription plans, Anthropic is rolling out access in stages. From launch through June 22, 2026, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost. After June 22, usage credits will be required. Anthropic says they aim to restore Fable 5 as a standard part of subscription plans as soon as capacity allows, and will communicate changes ahead of time.

Project Glasswing and Claude Mythos 5

Project Glasswing is Anthropic's initiative — developed in collaboration with the US government — to deploy Mythos-class AI to trusted cyberdefenders and critical infrastructure providers. Claude Mythos 5 is the latest upgrade deployed through this program, succeeding Claude Mythos Preview.

The program has already shown results: Project Glasswing partners used earlier Mythos models to help secure critically important software systems. Access to Mythos 5 is expanding from initial US government partners to hundreds of organizations across 15 countries, with a focus on infrastructure operators.

Anthropic has been clear that the dual-use nature of Mythos-level capabilities is exactly why public access is gated. The same skills that help a defender find vulnerabilities could assist an attacker exploiting them. That tension drives the Fable 5 vs. Mythos 5 distinction — same model, different guardrails.

Frequently Asked Questions

What This Means for the AI Landscape

Claude Fable 5 is a genuine step forward, not a minor revision. The 50-million-line codebase migration at Stripe is the kind of real-world result that reframes how developers think about what AI agents can do. The Pokémon example is fun, but what it demonstrates — autonomous operation over long tasks with minimal scaffolding — is what matters for production use.

Anthropic is also threading a difficult needle here. Mythos-level capabilities are real and dual-use. The company's approach — restrict the most dangerous configurations to vetted partners, release a safety-filtered version publicly — is a considered one. Whether the classifiers are calibrated right is something developers will discover quickly in practice.

For anyone building on Claude today, the transition to Fable 5 requires updating how you handle API responses. But the performance gains — especially on complex, long-horizon coding tasks — make that update worth planning for now rather than later.