Claude Fable 5 Brings Mythos-Class AI to Developers

Anthropic released Claude Fable 5 on June 9, 2026, making its first Mythos-class model broadly available after previously limiting Mythos-level capability to vetted cyberdefense partners. The company says Fable 5 is more capable than any Claude model it has generally released before, especially on long, complex software engineering, knowledge work, vision, and scientific tasks.

The launch is not just another Opus upgrade. Anthropic now describes Mythos-class models as a tier above Opus. Fable 5 is the public version with conservative safety classifiers, while Claude Mythos 5 uses the same underlying model with some safeguards lifted for approved users in Project Glasswing.

What changed

Fable 5 is aimed at long-horizon work where the model has to stay coherent across large projects rather than answer a single prompt well. In its launch post, Anthropic highlights software migration, finance analysis, high-resolution visual reasoning, long-context memory, and scientific research workflows as the areas where the model pulls away from previous Claude releases.

For developers, the platform details matter more than the launch-day benchmark claims:

Model	API ID	Availability	Context	Max output	Price
Claude Fable 5	claude-fable-5	Generally available	1M tokens	128k tokens	$10 / $50 per MTok input/output
Claude Mythos 5	claude-mythos-5	Limited Project Glasswing access	1M tokens	128k tokens	$10 / $50 per MTok input/output

Anthropic’s developer docs say Fable 5 is available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Mythos 5 is not self-serve; approved customers must go through Anthropic, AWS, or Google Cloud account teams.

The price is double Claude Opus 4.8, which is listed at $5 per million input tokens and $25 per million output tokens. That means Fable 5 should be treated as a top-tier tool for hard tasks, not a default routing choice for every request. If you are already tracking LLM API costs, this belongs in the same budget class as high-autonomy agent runs and expensive evaluation loops.

The safety architecture

The defining product decision is fallback. Fable 5 includes classifiers for sensitive areas including cybersecurity, biology and chemistry, and model distillation. When those classifiers trigger, Anthropic says the request is handled by Claude Opus 4.8 instead of Fable 5, and users are told when that happens.

Anthropic says the safeguards trigger in less than 5% of sessions on average, meaning more than 95% of sessions use Fable 5 throughout. The company also says the safeguards are tuned conservatively at launch, so some harmless requests will be caught while it narrows false positives.

This matters for production integrations because “model selection” is no longer a simple provider-side constant. With Fable 5, your system may receive a refusal or a fallback response depending on the request domain. The Claude API docs say refused requests return HTTP 200 with stop_reason: "refusal" rather than an API error. They also document fallback options and say customers are not billed for a request that is refused before output is generated.

If you run coding agents, vulnerability scanners, biology workflows, or model-evaluation pipelines, test your actual prompts before swapping this into a live router. A benchmark win is less important than knowing when your workload crosses a classifier boundary.

Mythos 5 stays restricted

Claude Mythos 5 is the same underlying model as Fable 5, but with some safeguards removed. Anthropic says it will initially be deployed through Project Glasswing, its cyberdefense program built with infrastructure providers, security teams, and government partners.

That distinction is important. Fable 5 is the general-use product. Mythos 5 is the controlled-access version for organizations Anthropic believes need stronger cybersecurity or, later, biology capabilities under a trusted access program. On June 2, Anthropic said Project Glasswing was expanding from roughly 50 initial partners to about 150 more organizations across more than 15 countries, after participants found more than 10,000 high- or critical-severity security flaws.

Data retention changes

Fable 5 and Mythos 5 also introduce a new data-retention policy for this capability tier. Anthropic’s Help Center says prompts and outputs for Mythos-class models are retained for 30 days for trust and safety purposes on every platform where the models are offered.

The practical effect is strongest for organizations using zero data retention. Anthropic says covered Mythos-class models are not available under ZDR, and organizations may need to enable retention or use a separate workspace, subscription, or cloud configuration before accessing them.

For enterprise teams, this is now part of the model-selection checklist:

• Can the workload tolerate 30-day retention?
• Does the workload involve domains likely to trigger fallback?
• Is the cost justified by fewer agent steps, better long-context reliability, or stronger visual reasoning?
• Do evals measure fallback behavior as well as normal completions?

That last point is easy to miss. If you are building AI agents, a hidden refusal path can create strange behavior several tool calls later. Add explicit tests for stop_reason: "refusal" and route those cases deliberately.

What to do next

Fable 5 is the first broad test of a model release pattern we should expect to see more often: frontier capability exposed through a public model, with the riskiest slices gated by classifiers, fallback models, retention, and trusted-access programs.

Use it where the task really benefits from long-context autonomy: multi-file codebase work, deep document analysis, high-resolution screenshots, complex spreadsheet reasoning, or research workflows that require sustained planning. Keep cheaper Claude models on routine traffic, and build telemetry that records when Fable 5 falls back or refuses.

The headline is that Mythos-level capability is no longer only a private preview. The engineering takeaway is narrower and more useful: advanced model routing now includes safety policy, retention policy, and fallback semantics as first-class production concerns.