Anthropic Opens Mythos-Class AI to the Public With Guardrails

Stripe's engineers spent one day doing what their full team estimated would take more than two months: a codebase-wide migration across a 50-million-line Ruby repository. That benchmark, from early access testing, marks the public debut of Claude Fable 5, which Anthropic released on June 9.

Fable 5 is Anthropic's first Mythos-class model available for general use. The Mythos tier had been restricted since its preview to a vetted circle of cyberdefenders through Project Glasswing, after Anthropic concluded the risk profile was too high for open deployment. Fable 5 is the company's resolution to that tension: the same underlying intelligence wrapped in a new layer of automated safety routing, with no change to the base model.

According to MacRumors, the model outperforms Anthropic's Opus line on longer, more complex tasks. Fable 5 can also sustain autonomous operation over longer horizons than any prior Claude release, making it the company's most capable model for multi-step workflows.

The safeguard logic

Queries touching cybersecurity, biology, chemistry, and model distillation are automatically rerouted to Claude Opus 4.8 rather than answered by Fable 5 directly. Anthropic projects this fallback will trigger in fewer than five percent of sessions on average. The approach has already drawn criticism: Price Per Token flagged reporting that cybersecurity researchers are unhappy with the guardrails, arguing the restrictions undermine legitimate professional workflows.

For the majority of use cases, the routing does not engage. IBTimes reports that on Hebbia's Finance Benchmark, which evaluates senior-level financial reasoning, Fable 5 scored higher than any model previously tested. Trading firm IMC validated strong performance across factual lookup, root-cause analysis, and expected-value analysis. On vision tasks, the model can reconstruct a web application's source code from screenshots alone and navigated Pokémon FireRed to completion using only raw game screenshots.

Mythos 5: the restricted layer

Claude Mythos 5, released alongside Fable 5, uses the same base model with some safeguards removed. Access is limited to Project Glasswing partners and selected biology researchers; Apple is among the named partners. According to Yahoo Finance, Mythos 5 achieved roughly 10x acceleration in parts of a drug-design pipeline and generated novel scientific hypotheses preferred by human researchers around 80% of the time. A week-long autonomous genomics research task saw it outperform a recently published specialized model despite being 100 times smaller in parameter count.

Anthropic describes Mythos 5 as carrying the strongest cybersecurity capabilities of any publicly acknowledged model. Access is expected to expand through a broader trusted access program, with no timeline given.

Pricing and access

Fable 5 costs $10 per million input tokens and $50 per million output tokens, less than half the price of the Mythos Preview it supersedes. The model is included in Pro, Max, Team, and seat-based Enterprise plans through June 22; from June 23 onward, usage credits are required until Anthropic can scale capacity. Business users of Mythos-class models face mandatory 30-day data retention for safety monitoring, a compliance consideration that will matter for enterprise deployments handling sensitive data.

What this architecture means

The tiered release reflects a debate the artificial intelligence research community has been navigating for years: whether graduated access with capability-matched oversight outperforms blanket restriction. Anthropic's approach here automates that oversight at inference time rather than at the access layer. Instead of limiting who reaches the model, it limits what the model will say, dynamically, per query. That is a meaningful architectural choice, not only a policy one.

Known risk: averages mislead. A five-percent fallback rate across all sessions could translate to substantially higher rates for security researchers, biologists, and chemists, the professionals who most need Mythos-class performance. If the routing is conservative enough to frustrate legitimate work at scale, pressure to circumvent it will grow.

Anthropic has bet that a public Mythos-class model with inline safeguards is safer than keeping the capability locked to a small cohort. The company has staked part of its safety reputation on getting that calibration right. Whether the routing logic satisfies both safety goals and practitioner needs is the question the next few months of adoption will answer.

FAQ

What is Claude Fable 5 and how does it differ from Claude Opus?
What topics trigger the automatic fallback to Claude Opus 4.8?
Who can access Claude Mythos 5 through Project Glasswing?
How is Claude Fable 5 priced and when does free subscription access end?