Anthropic Is Reportedly Building Two Models That Could Redefine AI. One of Them Has 10 Trillion Parameters.

According to insider sources, Anthropic is quietly preparing to launch two of the most ambitious AI models the industry has ever seen: Mythos and Capybara. If the leaked details are accurate, both models represent a massive leap beyond anything currently available, including Anthropic's own flagship Claude Opus 4.6.

What We Know About Mythos

Sources familiar with the project claim Mythos will significantly outperform Claude Opus 4.6 across virtually every benchmark, with particular dominance in coding and complex multi-step reasoning tasks.

But the detail generating the most concern in security circles is this: Mythos can reportedly identify and exploit software vulnerabilities faster than human teams can patch them. That capability alone has raised alarm bells. If true, it would make Mythos one of the most powerful offensive cybersecurity tools ever created, and Anthropic knows it.

Capybara: The Most Expensive Model Ever Built

While Mythos focuses on raw intelligence, Capybara appears to be an entirely different beast. Sources indicate the model will feature approximately 10 trillion parameters, which would make it, by a significant margin, the largest and most expensive neural network ever trained.

For context, GPT-4 was estimated at roughly 1.7 trillion parameters. Capybara would represent nearly a 6x increase over that, requiring computational resources that only a handful of organizations on the planet could even attempt to provision.

The training costs alone would dwarf anything the industry has seen. Current estimates for training frontier models sit around $100-200 million. A 10-trillion-parameter model could push that figure into the billions.

Why Both Models Are Currently Locked Down

Here is where the story gets interesting. According to insiders, both Mythos and Capybara are functionally complete. They exist. They work. But Anthropic has chosen not to release them publicly.

The stated reason: "unprecedented cybersecurity risks."

Currently, access to both models is reportedly limited exclusively to information security specialists who are evaluating the models' capabilities and potential for misuse. This is consistent with Anthropic's Responsible Scaling Policy, which requires safety evaluations before any model deployment.

The implication is clear: Anthropic believes these models are too capable to release without extensive safeguards. In an industry racing to ship products faster than competitors, that is an extraordinary decision.

Anthropic Is Making Its Own Flagship Obsolete

There is a certain irony in the situation. Anthropic built its reputation on Claude, particularly the Opus line, which many developers consider the gold standard for complex reasoning tasks. Now the company has apparently built something that makes its own best product look outdated.

The gap described by insiders between Opus 4.6 and Mythos is reportedly not incremental. It is, according to one source, "a generational leap."

If Anthropic eventually releases Mythos, it would effectively cannibalize its entire existing product line. Every enterprise customer using Claude Opus would want to upgrade immediately.

What This Means for the AI Race

The parameter race is not over. While much of the industry has shifted focus toward efficiency and smaller models, Capybara suggests that raw scale still produces capabilities that cannot be achieved through optimization alone.

Cybersecurity is becoming the deployment bottleneck. The fact that both models are being held back specifically due to security concerns signals a fundamental shift. We may be entering an era where the hardest problem in AI is not building capable models but deciding whether to release them.

Anthropic's safety-first positioning is being tested. The company has long argued that building frontier models responsibly is better than not building them at all. Mythos and Capybara are the ultimate test of that philosophy.

The competitive dynamics are about to shift. If Anthropic has models this powerful waiting in reserve, it changes the calculus for OpenAI, Google DeepMind, and every other frontier lab.

Our Take

We should be clear: none of this is officially confirmed. Anthropic has not publicly acknowledged either Mythos or Capybara. Everything described here comes from insider sources and should be treated accordingly.

But the pattern is consistent with what we know about Anthropic's research trajectory. The company has been hiring aggressively, securing massive funding rounds, and publicly discussing the need for pre-deployment safety testing.

The real question is not whether these models exist. It is what happens when they, or something like them, eventually become available. Because if Anthropic has built this, others are not far behind. And not everyone will choose to keep the door locked.