Anthropic's Claude Mythos Leaks: What We Know So Far

An unsecured database exposed roughly 3,000 unpublished assets from Anthropic's blog last Thursday evening. Among them: a draft announcement for Claude Mythos, which the company calls "a step change" in performance and "the most capable we've built to date."

The leak itself tells you something about the current state of AI development—moving so fast that even security basics get overlooked in the rush to ship. But what's more interesting is what Anthropic chose to say in that unfinished blog post, and what they're choosing to do now that it's public.

The Numbers and the Caution

Mythos represents a new tier above Opus, Anthropic's previous flagship. The leaked post claims "dramatically higher scores" on software coding, academic reasoning, and cybersecurity benchmarks compared to Claude Opus 4.6. Those are the same categories every AI lab has been racing to improve since GPT-4 set the baseline.

What makes this announcement unusual is the explicit hand-wringing. Anthropic writes that Mythos is "a large compute intensive model" that will be "very expensive for us to serve and will be very expensive for our customers to use." They're working to make it more efficient before general release.

More telling: they're starting with "a small number of early access customers" specifically to explore cybersecurity applications and risks. The company wants to "understand the model's potential near-term risks in the realm of cyber security and share the results to help cyber defenders prepare."

I've covered enough AI releases to recognize when a company is genuinely worried versus when they're performing concern for PR purposes. This reads like genuine worry.

The Name Game

The choice of "Mythos" has generated predictable commentary. Many in the AI safety community immediately connected it to H.P. Lovecraft's Cthulhu Mythos—stories about unknowable cosmic horrors that drive people mad. One observer noted the irony: "Anthropic's mysterious spooky new model is codenamed mythos while OpenAI named theirs after a freaking potato."

Anthropic's explanation is more prosaic. They chose the name "to evoke the deep connective tissue that links together knowledge and ideas." Whether that clarification makes the choice better or worse depends on your tolerance for corporate mysticism.

The model may have also been code-named "Capybara" internally, though that detail remains unconfirmed. Fortune reports that the cache contained close to 3,000 unpublished assets, suggesting this leak exposed far more than just one model announcement.

The Pattern of Caution

Anthropic has built its brand on being the safety-conscious AI company. They've published extensive research on constitutional AI, mechanistic interpretability, and model behavior. This cautious rollout fits that pattern.

But caution also serves business interests. A gradual release to select customers creates exclusivity and lets Anthropic control the narrative. Early access means early feedback from customers who pay premium prices and have real-world use cases. It's good practice dressed up as good citizenship.

The cybersecurity focus is particularly interesting. If Mythos significantly outperforms existing models at finding vulnerabilities or writing exploits, releasing it widely could create immediate problems. But if it also excels at defending against those same threats, cyber defenders need access to it quickly. Anthropic appears to be trying to thread that needle by releasing to security-focused customers first.

The leaked post doesn't specify a timeline for broader availability. Given the emphasis on making the model more efficient and understanding its risks, this could be months away. Or it could be weeks—the draft was clearly unfinished, and plans change.

What This Means for the Race

The leak comes at a moment of intense competition. OpenAI recently upgraded its Codex tool with plugins, explicitly taking a shot at Anthropic over Claude's recent usage limit reductions. Google released Gemini 3.1 Flash Live, which promises more natural voice interactions. The technical gaps between frontier models keep narrowing.

Anthropic is reportedly considering an IPO as soon as October, with Bloomberg suggesting they want to go public before OpenAI. That timeline would make the Mythos release strategically important—going public with your most capable model still in limited access doesn't send the strongest signal to investors.

The computational cost that Anthropic emphasizes is worth noting. We're entering a phase where the most capable models are also the most expensive to run. That creates natural segmentation in the market: smaller, faster models for everyday tasks, larger models for specialized work where the extra capability justifies the cost. Anthropic's acknowledgment that Mythos is expensive suggests they're already thinking about where it fits in that landscape.

The Unsolved Problem

Here's what the leak doesn't answer: what specific capabilities make Mythos a "step change" beyond better benchmark scores? AI labs have learned to game benchmarks. What matters is whether the model can do meaningfully new things.

The cybersecurity focus hints at one possibility—that Mythos crosses some threshold in its ability to find and exploit vulnerabilities autonomously. That would certainly justify the cautious rollout. But it could also mean the model is simply better at everything in ways that create second-order risks Anthropic hasn't fully mapped yet.

The draft blog post was, by Anthropic's own characterization, "very undercooked." It ended abruptly, suggesting they were still figuring out how to message this release. Now that the information is public through a leak rather than a controlled announcement, they're in reactive mode.

For those keeping score on AI progress, Mythos represents another data point in the same direction: capabilities keep improving, costs keep rising, and the gap between what these systems can do and our understanding of their risks keeps widening. Whether Anthropic's cautious approach narrows that gap or just delays the inevitable remains to be seen.

Bob Reynolds is a Senior Technology Correspondent for Buzzrag.