Anthropic is testing Mythos 5.0, a model so potent it scares them

When an AI lab trains a model that shatters its own internal benchmarks, you would expect a splashy keynote. Instead, Anthropic kept its newest beast locked away behind closed doors, genuinely unnerved by what it can do.

Anthropic has quietly begun testing its next-generation flagship model, Claude Mythos 5.0, with a select group of early-access customers. The rollout comes hot on the heels of a massive slip-up: human error in configuring their content management system accidentally exposed an unpublished blog draft and nearly 3,000 internal assets to an open, searchable data store, where cybersecurity researchers found them.

A tier above Opus

The leaked documents describe Claude Mythos, also referenced under the tier name Capybara, as an entirely new class of AI sitting above Opus. Historically, Anthropic segmented its family into Haiku, Sonnet, and the heavyweight Opus. Mythos breaks that ceiling.

Compared to Claude Opus 4.6, Anthropic notes that this new model delivers a dramatic step change across software engineering, academic reasoning, and offensive cybersecurity. Just 48 hours after the leak made headlines, developers spotted Mythos 5.0 (Beta) appearing inside the Claude interface, while the Claude Code terminal officially labeled it as Anthropic's next-generation model.

That raw power comes at a steep price. The leaked draft explicitly notes that the model is prohibitively expensive to run and nowhere near ready for a broad public release.

Outpacing cyber defenders

The real reason Anthropic is dragging its feet comes down to cyber risk.

Anthropic admitted in the draft that Mythos is far ahead of any rival system in cyber capabilities. In the wrong hands, it could automate and amplify sophisticated network attacks at a scale existing cyber defenses cannot handle. The company is treading carefully, hoping to brief and prep defensive teams before unleashing it.

Inside Anthropic, this level of intelligence is already restructuring how engineers build software. While CEO Dario Amodei previously projected AI writing most code within a year, internal teams are already there. An engineer who recently joined revealed that his team hasn't written code by hand for months. Instead, they operate as product managers directing swarms of parallel agents.

Why it matters

For everyday developers, you probably won't be prompting Mythos 5.0 anytime soon. Its immense inference costs mean Anthropic will keep it cordoned off for deep-pocketed enterprise pilot programs. But it sends a crystal-clear signal: coding is evolving from AI-assisted typing into multi-agent orchestration, and the barrier between building software and directing software has melted away.

For the broader industry, the leak highlights a terrifying asymmetry in AI safety. Network offense is notoriously cheaper and faster than network defense. Releasing a model with unprecedented exploit capabilities could easily spark catastrophic zero-day waves, inviting severe regulatory backlash across the entire frontier AI sector.

We think Anthropic’s extreme caution here is both genuine and self-preserving. Rushing Mythos out to claim benchmark bragging rights would be reckless. Until defensive toolkits can digest the firepower of this new tier, moving slowly isn't just safety theater; it is the only way frontier labs avoid blowing up their own runway.