Anthropic Announces Claude Fable 5 and Mythos 5 - Mythos-Class Models Go Public

On June 9, 2026, Anthropic simultaneously announced two new AI models: Claude Fable 5 and Claude Mythos 5¹. Both are the same underlying model — Fable 5 is the generally available version with safeguards enabled, while Mythos 5 is a restricted version with safeguards lifted in certain areas¹³. Anthropic calls this new model tier “Mythos-class,” positioning it above the existing Opus family.

According to Anthropic, Fable 5’s capabilities exceed those of any model the company has ever made generally available, achieving state-of-the-art results on nearly all tested AI capability benchmarks¹. Pricing is set at $10 per million input tokens and $50 per million output tokens — less than half the price of the previously limited-access Claude Mythos Preview¹.

Fable 5’s Capabilities - From Software Engineering to Vision

Among the examples presented in the announcement, the software engineering results stand out. In testing at payments company Stripe, Fable 5 compressed months of engineering work into days, completing a migration on a 50-million-line Ruby codebase in one day that would normally take two months¹. In code evaluation, it achieved the highest score among frontier models on Cognition’s FrontierCode benchmark, even at a medium effort setting¹.

Vision (image understanding) capabilities have also been strengthened. The model can extract accurate figures from scientific diagrams and reconstruct the source code of a web app from screenshots alone¹. As a demonstration, it fully completed Pokémon FireRed using only the game screen as input¹. AWS’s introduction article also highlights its understanding of diagrams, charts, and tables in PDFs, along with proactive self-verification that updates its own skills based on what it learns².

Long-running task handling has advanced as well. Fable 5 reportedly maintains focus on long-horizon tasks spanning millions of tokens and uses its own notes to improve its output¹. On the technical side, the context window is reported at 1 million tokens with a maximum output of 128,000 tokens per response³. The API model ID is claude-fable-5³.

How the Safeguards Work - Opus 4.8 Answers When Triggered

What distinguishes Fable 5 is not only its capabilities but also the design of its safety measures. Classifier-based safeguards are implemented in three areas: cybersecurity, biology and chemistry, and distillation (capability extraction aimed at developing competing models)¹.

When a request in a covered area is detected, the system uses a fallback mechanism in which Claude Opus 4.8 automatically provides the response. According to Anthropic, more than 95% of Fable sessions involve no fallback at all¹. Per AWS, when a request is routed to Opus 4.8, customers pay only Opus pricing².

Robustness was also verified externally. An external bug bounty found no universal jailbreak across more than 1,000 hours of testing, and one external partner’s evaluation reported zero compliance with harmful single-turn requests even when using 30 different publicly known jailbreak techniques¹. In addition, Mythos-class models require 30-day data retention on all traffic, and Anthropic states the retained data will not be used to train new Claude models or for purposes other than safety¹.

Mythos 5 and Project Glasswing - Limited Availability for Cyberdefense

Mythos 5, meanwhile, is the same underlying model as Fable 5 but with cybersecurity safeguards lifted¹. Access is limited to cyberdefense organizations partnering through Project Glasswing, a program run in collaboration with the US government¹. Anthropic released its first Mythos-class model, Claude Mythos Preview, through the same program to cyberdefenders and critical infrastructure providers in April 2026, and Mythos 5 is the upgrade to that model¹.

Mythos 5 is said to have the strongest cybersecurity capabilities of any model in the world¹. Results have also been shown in the life sciences: Anthropic’s internal protein design experts used Mythos 5 to accelerate parts of the therapeutic development process by roughly ten times, generating drug design candidates for 9 of 14 protein targets¹. While most biology and chemistry requests currently fall back to Opus 4.8, a trusted access program for biology researchers is expected to launch soon¹.

Availability and Pricing

Fable 5 became immediately available via the API and usage-based enterprise plans on launch day¹. For subscription plans, a phased rollout is planned: through June 22, it is included at no extra charge in Pro, Max, Team, and per-seat Enterprise plans; from June 23, usage credits will be required. Anthropic plans to gradually return it to standard availability as capacity allows¹.

On AWS, the model is available through Amazon Bedrock (US East, Europe Stockholm) and Claude Platform on AWS (North America, South America, Europe, and Asia Pacific)². The $10 input / $50 output pricing (per million tokens) is exactly twice that of Opus 4.8³.

What the “One Model, Two Tiers” Release Approach Means

Beyond the performance gains themselves, what deserves attention in this announcement is the release approach: offering the same model under two names, distinguished by whether safeguards are enabled. In dual-use domains like cybersecurity and the life sciences — where advanced capabilities serve both attack and defense — uniformly restricting capability would also block legitimate researchers and defense organizations. Protecting general users with classifiers and fallbacks while providing an unrestricted model to vetted organizations can be seen as one practical answer to the capability-safety trade-off.

On the alignment front, automated evaluations of Mythos 5 found levels of misaligned behavior as low as Opus 4.8, and since Fable 5 shares the same underlying model, it is described as having the same property¹. With top-tier capabilities now in the hands of general users, this release may also influence how other companies structure their future model offerings.

Detailed technical specifications and API usage information are available in Anthropic’s official documentation and AWS’s introduction article.