Amodei says Anthropic’s safety delay may have cost it an AI lead

Anthropic’s CEO is admitting the obvious tradeoff: slowing down can hurt you. In a lengthy X post on Tuesday, Dario Amodei said the company held back broader access to its Mythos-class models until June so it could tighten cybersecurity safeguards, even though Anthropic first released the models in February. He said that caution may have cost the company “a significant chunk of a several-month model lead.”

His defense is simple enough. Releasing the models sooner, he argues, would have given organizations less time to prepare for more capable AI-driven cyberattacks. Anthropic says the delay was in line with its Responsible Scaling Policy, which requires a strong case that catastrophic risk is contained and says development and deployment should be slowed as far as needed until the company no longer believes it has a meaningful lead.

The company’s policy story, though, has clearly become a fight. Nathan Calvin has accused Anthropic of looking less like a safety-first lab and more like a company chasing superintelligence while softening its own guardrails. He points to more than $500 billion in computing deals, the company’s growing use of AI to build new models, and changes to the Responsible Scaling Policy.

Anthropic’s revised policy, version 3.0, dropped an older 2023 pledge never to train a model unless safety could be guaranteed in advance. Amodei and chief science officer Jared Kaplan say that earlier language was too rigid. Kaplan told Time in February that stopping training would not help if competitors were moving faster, and Amodei now frames the company’s approach as a “race to the top.”

That race has gotten complicated. Anthropic rolled out Claude Mythos Preview in April to selected partners, expanded access in June, then released Claude Fable 5 on June 9 before a U.S. export-control directive forced it to suspend both models three days later. The company later brought them back, and its newer access rules now divide cybersecurity use into tiers, with the most restrictive tier blocking Claude Opus 5.5 in 46 of 50 trials on one benchmark.

Amodei also says the company is trying to outrun a future where the safeguards lag behind model capability. Anthropic’s June report says Claude wrote more than 80% of the code merged into its codebase as of May, and the company has separately said that figure rises to 90% or more when scripts and experimental code are included. In that light, the delay looks less like hesitation and more like a lab trying to keep its own paper bag over a growing fire.