Three days after Anthropic launched its most capable model ever, they pulled it. Not because it broke, but because the U.S. government said so.
Claude Fable 5 went live on June 9, 2026, Anthropic’s first “Mythos-class” model made available to the public, a full tier above what had previously been its best generally available offering. On June 12, Anthropic suspended access globally to comply with a federal export control directive. It came back on July 1, with tighter safeguards attached. Same model, different guardrails, three-week gap in between.
That’s not a normal product cycle. It’s a decent hint that the model itself is a bigger deal than the average release-notes email suggests.
What “most capable model ever” means
Anthropic’s own numbers: Fable 5 is the first of its models to break 90% on its internal benchmark for complex, long-running analytical tasks, a 10-point jump over the model it replaced. It’s the highest scorer on an external frontier coding eval. On a hard physics research benchmark, Anthropic says it reached in 36 hours roughly what a competitor’s model took four days to reach.
The headline capability isn’t “smarter answers.” It’s duration. Fable 5 is built to run a multi-stage task: plan it, execute it, check its own output against the original goal, keep going, for extended stretches without someone checking in on every step. Previous-generation models needed a person steering constantly on anything past a narrow, well-defined task. This one is designed to be handed a messy, multi-day project and left alone.
The part Anthropic built in on purpose: it says no
Here’s the trade-off. A model this capable in, say, cybersecurity is also a model that’s dangerously good at finding and exploiting software vulnerabilities. So Fable 5 ships with classifiers that quietly reroute certain requests (cybersecurity, biology, chemistry, model distillation) to a different, less specialized model instead of answering directly.
Anthropic says this triggers in under 5% of real sessions. Fine, but read that the other way: on purpose, out of the box, this model is built to sometimes refuse you and hand you off to something weaker, without necessarily being right about when that’s actually warranted. Anthropic itself admits the safeguards were tuned conservatively and will catch some harmless requests along with the real ones. If your use case brushes up against any of those categories, expect friction, at least for now.
There’s also a version without those limits, Claude Mythos 5, but it’s not generally available. It’s restricted to a small number of vetted partners, mostly cyber defense and infrastructure work, through a government collaboration. The public doesn’t get the unrestricted version, on purpose, indefinitely.
Where Fable 5 shines and where it’s just expensive
Fable 5 costs $10 per million input tokens and $50 per million output tokens, roughly double the rate of Anthropic’s next tier down. The gap in ability grows with the length and complexity of the task. On something quick and well-scoped, it isn’t meaningfully better than a cheaper model, and the price difference is pure waste.
Where it earns its cost: multi-day engineering work, deep research synthesis, anything document-heavy where the model needs to read diagrams, tables, and charts buried in files and reason across all of it coherently. If your business runs on quick, scoped AI tasks (a support macro, a summarization step, a single-turn classification) this model is the wrong tool and the wrong price for the job.
What´s next?
Every capability jump like this quietly moves the line on what’s worth automating. A process that wasn’t a sane candidate for AI six months ago might be one now, purely because the model changed underneath it — nothing about your business had to change.
The mistake is assuming that means you should hand more to AI by default. The Fable 5 launch itself is a decent argument for the opposite: even Anthropic, the company that built it, didn’t fully trust its own safeguards enough to leave it running without a three-week pause to tighten them. That’s not a knock on the model. It’s a reminder that “most capable” and “safe to deploy without oversight” aren’t the same claim, even from the people who made it.
If you’re trying to figure out whether a model like this actually changes anything for your operations: Anthropic wasn’t sure enough to skip the three-week pause. You probably shouldn’t be either, at least not without checking what this fixes for you first.