Anthropic says holding back its Mythos models for safety likely cost it its lead

A delay Anthropic now admits was costly

Anthropic held back broader access to its Mythos-class AI models for about four months, from February 2026 until June 2026. As TechGig reported, the reason given by CEO Dario Amodei was cybersecurity: the company wanted stronger safeguards in place first. Amodei also conceded that the wait may have cost Anthropic a significant competitive edge.

He laid this out on X, in reply to criticism from Nathan Calvin about how the company handles AI safety. His argument was about timing. Had the models gone out sooner, organisations would have had less time to get ready for cyberattacks that use increasingly capable AI.

A "Mythos-class" model simply means a model in the same top tier of capability as Anthropic's Mythos line. The company has not framed the delay as a mistake, only as a trade it chose to make.

How the rollout actually unfolded

Anthropic did not keep Mythos locked away entirely. In April it opened Claude Mythos Preview to selected partners through an effort called Project Glasswing. Those partners used it for defensive work, meaning protecting systems, while ordinary developers were left out.

On June 2 the program grew to about 150 organisations in more than 15 countries. A week later, on June 9, Anthropic released Claude Fable 5, a Mythos-class model meant for general use with extra safeguards added.

Then came a setback that had nothing to do with the company's own caution. A US government export-control directive (a rule limiting where advanced technology can be sent) forced Anthropic to suspend Fable 5 and Mythos 5 three days after the launch. The Commerce Department lifted those controls on June 30.

Safety rules that were rewritten

Anthropic's Responsible Scaling Policy is its public rulebook for handling dangerous AI capabilities. It says the company will work to contain catastrophic risk and will hold back deployment when needed. In February 2026 the company issued version 3.0, which replaced the stricter, fixed requirements of earlier editions.

The newer approach leans on openness: publishing information about risks and inviting outside experts to evaluate its models. The policy has since been updated again, and version 3.4 took effect on July 8.

One change drew attention. The revision dropped a promise made in 2023 to have safety measures guaranteed ready before a model is even trained. Chief science officer Jared Kaplan told Time that going it alone on such pledges made little sense when rivals were "blazing ahead." Amodei's counter-case is that staying at the cutting edge gives Anthropic more influence over industry safety norms than standing on the sidelines would.

AI that helps build AI

The story has a second thread that explains the urgency. In a June report titled "When AI builds itself," Anthropic said Claude wrote over 80 per cent of the code merged into the company's own codebase as of May. In other words, its AI is already doing most of the programming work on its own successors.

Amodei acknowledged that AI progress could soon move faster than the safeguards meant to keep it in check. That is the heart of the worry behind the delay: if capability accelerates, the gap between what models can do and what defenders are ready for can widen quickly.

What Anthropic has promised next

The company has set two dates in its Frontier Safety Roadmap. By December 15, 2026, it aims to put forward a policy proposal on how governments could support evaluators embedded inside AI companies, meaning outside checkers who work close to the models.

By January 15, 2027, it plans to publish a framework for pacing AI development. If it chooses not to, it says it will explain why. Those deadlines give readers something concrete to hold the company to.

TechGig dated its report October 23, 2026. Our source does not say how far behind, if at all, Anthropic now sits against rivals, so the size of the lost lead remains an open question.