Andrew Curran says a stronger Anthropic Mythos model has emerged from training

Why it matters

If the claim is accurate, Anthropic may already be beyond the models Washington just restricted, exposing a gap between frontier-model iteration and model-specific export controls.

Andrew Curran (@AndrewCurran_) said Sunday that a more capable version of Anthropic's Mythos-class model has come out of training, a claim that has not been confirmed by Anthropic and lands nine days after the US government forced Anthropic to disable Fable 5 and Mythos 5.

In a 17-post thread on X, Curran wrote that he did not know whether the model would be called "Mythos 5.1" or "Mythos 6," or whether Anthropic would keep it internal to speed further development. He added in replies that the original Mythos model finished training on February 7 and that the current government restrictions apply only to Fable 5 and Mythos 5, not to an unnamed successor.

The important word is "claim." Anthropic has not announced a new Mythos model, has not updated its public model pages with a successor, and has not said whether a new training run has completed. Curran, whose account regularly posts on frontier AI model development, told one respondent that he did not know anything more, and told another that this post was not merely a guess. That is not the same thing as confirmation from Anthropic.

But the timing makes the claim consequential even before it is verified. Anthropic, co-founded and led by Dario Amodei, has spent the last two weeks in a public fight over who gets to decide when a frontier model is too capable to ship. On June 9, Anthropic launched Claude Fable 5 and Claude Mythos 5, describing Fable 5 as a Mythos-class model made safe for broader use and Mythos 5 as the same underlying model with some safeguards lifted for selected cyberdefense and infrastructure partners. Anthropic said Mythos-class models sit above its Opus class in capability, and priced Fable 5 and Mythos 5 at $10 per million input tokens and $50 per million output tokens.

Three days later, the rollout broke. In a June 12 statement, Anthropic said the US government had issued an export-control directive requiring the company to suspend access to Fable 5 and Mythos 5 for every foreign national, including foreign national Anthropic employees inside the United States. Anthropic said the net effect was that it had to disable both models for all customers to ensure compliance.

Anthropic said the government had not given specific details of its national-security concern. The company said its understanding was that officials believed they had become aware of a way to bypass, or jailbreak, Fable 5. Anthropic disputed that the reported technique justified recalling the model, saying the demonstrated vulnerabilities appeared minor, previously known and discoverable by other public models without a bypass.

That fight turned a product launch into a governance test. Fable 5 was the version Anthropic was willing to expose broadly, with safeguards in areas including cybersecurity, biology and chemistry. Mythos 5 was the restricted version for approved partners. If Curran is right that a more capable Mythos successor has finished training, the question is no longer just whether Fable 5 and Mythos 5 should be restored. It is whether the control system Washington just used can keep up with model iteration when the order names specific models rather than a capability threshold.

The model-specific structure matters. Curran's thread explicitly points to that gap: the bans, as described publicly by Anthropic and in reporting, target Fable 5 and Mythos 5. They do not mention a future model called Mythos 5.1, Mythos 6 or anything else. Anthropic has not said whether it would treat a successor as covered by the same directive, seek separate clearance, or keep it internal.

Anthropic's own launch materials already framed Mythos as a moving target. On June 2, the company said it was expanding Project Glasswing to about 150 new organizations in more than 15 countries, and warned that within 6 to 12 months many other AI companies could have Mythos-class models. On June 9, Anthropic said more capable models were arriving in the coming months and that Fable 5's safeguards could trigger in less than 5% of sessions on average while still sometimes catching harmless requests.

The safety record is not clean enough for either side to claim an easy victory. A June 16 red-team study of Fable 5 and Opus 4.8 found that Fable 5 resisted most automated jailbreak attacks better than Opus 4.8, but still produced 702 panel-confirmed harmful completions across the tested harm categories. The study's strongest adaptive attack broke Fable 5 on 6.1% of intents in its worst case. That does not validate the government's reported move against Fable 5, but it does support the broader premise that even heavily tested frontier models retain a live attack surface.

The commercial incentives are just as visible. Fable 5 was Anthropic's path to turning Mythos capability into revenue across Claude API and enterprise usage. Mythos 5 was the path to controlled access for the customers Anthropic wanted to trust with stronger cyber capabilities. Pulling both offline cut directly against the model launch economics Anthropic had just announced. Keeping a successor internal, if one exists, would preserve Anthropic's research acceleration while avoiding another immediate fight over access. Shipping it would force the same confrontation back into the open.

That is why Curran's unverified post is being read as more than model gossip. Anthropic's June clash with Washington was not only about one jailbreak claim. It was about whether frontier AI can be regulated by emergency orders aimed at named products while labs continue to train successors on a faster clock. If a stronger Mythos-class model has indeed finished training, Anthropic's next decision will show whether the new bottleneck in AI deployment is compute, safety engineering, customer demand or permission from the state.