Anthropic Leak Reveals “Mythos” — Codename ‘Capybara’ — Reportedly Far Ahead of Claude Opus 4.6

Leak and verification

Nearly 3,000 internal files from Anthropic were left exposed after a misconfigured content-management system, and it has been reported that a Cambridge security researcher, Alexandre Pauwels, validated the cache before Fortune and others combed through the materials. Anthropic confirmed the existence of the model described in the documents. The draft blog describes a new layer above the company’s long-assumed ceiling — Claude Opus 4.6 — giving the project the working name Mythos and the internal codename “Capybara.” How did a company that preaches “safety first” leave its roadmap in public? A forgotten privacy toggle, reportedly.

Capabilities, costs and an unusual release plan

The leaked evaluation reportedly shows Mythos outperforming Opus 4.6 by a wide margin on core benchmarks: software programming, academic reasoning (math, science and formal logic), and even offensive cybersecurity capabilities — the latter raising explicit, unusual alarms inside Anthropic’s drafts. The company has been quoted calling it a “qualitative leap” and “its most powerful model to date,” and it has been reported that Anthropic plans to give first access not to general developers but to defensive cybersecurity teams to let defenders get the tools before wider release. The catch: the draft admits running Mythos is “very expensive,” and Anthropic says it will need substantial efficiency gains before any mass deployment.

Irony, brand clash and geopolitical stakes

The episode is rich in irony: the same basic permissions error that has tripped up cloud buckets elsewhere exposed a project that, if the internal assessments hold up, could change the competitive landscape among the likes of OpenAI, Google and large Chinese players. The codename collision is comic and pointed — Alibaba’s Tongyi Qianwen (通义千问) already uses capybara imagery, prompting a viral backlash and even jibes at industry figures: “Is Sam Altman asleep?” Beyond the memes, the leak lands amid growing international scrutiny of advanced AI and export-control debates; it has been reported that regulators in both Washington and Brussels will view such disclosures through a national-security lens. A CEO summit itinerary and guest list also surfaced in the files, underscoring how operational lapses can turn product strategy into a geopolitical story overnight.