Anthropic dropped two new models today, and this one feels different from the usual "we made it smarter" announcements. Claude Fable 5 and Claude Mythos 5 are the most capable models the company has ever made publicly available, and the way they've structured the release says a lot about where things are heading.
So What Are These Models?
At their core, Fable 5 and Mythos 5 are the same model. The difference is which safety guardrails are active.
Claude Fable 5 is the version anyone can use starting today. It has safety classifiers running in the background that will quietly reroute certain sensitive requests (mainly cybersecurity and biology topics) to Claude Opus 4.8 instead. Anthropic says this only happens in less than 5% of sessions, so for most people, most of the time, you're getting the full Fable 5 experience.
Claude Mythos 5 is Fable 5 with those cybersecurity restrictions lifted. Right now it's only available to a small group of vetted partners, mostly cyber defenders and critical infrastructure operators, through Anthropic's Project Glasswing in collaboration with the US Government.
The Capabilities Are a Genuine Step Up
Fable 5 tops the charts on nearly every benchmark Anthropic ran. A few things stand out:
Coding. Stripe ran a codebase-wide migration on a 50-million-line Ruby project using Fable 5. It finished in a day. Their estimate for doing it by hand? Two months with a full team. On Cognition's FrontierCode benchmark for agentic coding quality, Fable 5 ranks highest among all frontier models.
Finance and knowledge work. Fable 5 leads Hebbia's Finance Benchmark for senior-level reasoning. IMC, a trading firm, said it "aced" their evaluation suite across factual lookup, root-cause analysis, and expected-value reasoning. One firm noted it's the first model to break 90% on their complex analytics benchmark, a 10-point jump over Opus.
Vision. The model can rebuild a web app's source code from screenshots alone. It beat Pokémon FireRed using only raw game screenshots, with no maps, no navigation aids, nothing. Earlier Claude models needed a whole helper harness to attempt that. Fable 5 did it with vision alone.
Long-context and memory. Fable 5 stays focused across millions of tokens. In tests using the deck-building game Slay the Spire, giving the model persistent file-based memory improved its performance three times more than it did for Opus 4.8.
Drug design. Mythos 5 accelerated parts of the drug design process by roughly 10x for Anthropic's internal protein design team. The model chose binding sites, ran protein design tools, and recovered from its own mistakes without human help. Nine of 14 protein targets in the study yielded strong candidates currently under investigation.
Genomics research. Over more than a week of largely autonomous work, Mythos 5 assembled single-cell data across 138 animal species, then designed and trained a custom machine learning model to identify cells performing the same biological role across distantly related organisms. That model outperformed one published recently in Science, despite being 100 times smaller. Anthropic plans to publish the results.
Safety Is Really the Main Story
Until today, Mythos-class capabilities were only available to a limited group through Project Glasswing. The reason for that was simple: these models are powerful enough to provide real uplift to bad actors in cybersecurity and biology. Releasing them widely required safeguards strong enough to hold up against motivated, sophisticated attempts to circumvent them.
Anthropic says those safeguards are now ready, though they've deliberately tuned them to be strict rather than permissive. There are three main pieces:
Classifiers watch for requests related to cybersecurity, biology and chemistry, and "distillation" (attempts to extract Claude's capabilities to train competing models). When triggered, the request goes to Opus 4.8 instead of being refused outright. External red teamers spent over 1,000 hours trying to find a universal jailbreak and came up empty, though the UK AI Safety Institute made some early progress in a short initial testing window.
A 30-day data retention policy now applies to all Mythos-class traffic. Anthropic won't use it for training, but will use it to spot novel attacks playing out across multiple sessions. Everything gets deleted after 30 days in almost all cases.
Conservative tuning is a deliberate choice. The classifiers will catch some harmless requests, and Anthropic knows it. They've said upfront that reducing false positives is a priority as the system matures.
Pricing and How to Get Access
Both models are priced at $10 per million input tokens and $50 per million output tokens. That's less than half what Claude Mythos Preview costs.
If you're on a paid subscription plan (Pro, Max, Team, or Enterprise), Fable 5 is included at no extra cost through June 22. After that, it'll require usage credits until Anthropic has the capacity to bring it back as a standard feature. API users can access it now via claude-fable-5.
Mythos 5 stays restricted to Glasswing partners for the time being. A broader trusted access program for cybersecurity organizations and biomedical researchers is in the works.
The Bigger Picture
What's interesting about this launch isn't just the capability jump. It's the model for how Anthropic is releasing frontier AI going forward: tiered access based on how confident they are in the safeguards, with a general-availability version that degrades gracefully rather than refusing outright. Whether the classifier approach holds up at scale is the real open question. But the framing itself, that you can release something at the frontier if you're careful about how you do it, is worth watching.