Anthropic says AI risks are rising. Anthropic also built an internal model, reportedly called “Model 2,” that appears to be more powerful than its top-of-the-line Mythos. Those two facts sit uncomfortably next to each other, and the company’s answer to that tension is simple: keep the stronger model in-house and don’t release it.
I review AI toolkits for a living, which means I spend my days testing what companies actually ship, not what they claim to have in the lab. So when a major lab announces it has something better than its flagship product and then declines to sell it to me, I pay attention. That almost never happens in this industry. The default move is the opposite — ship the bigger model, charge more for it, and let the marketing team handle the risk questions.
What we actually know
The confirmed picture is narrow, so let me be honest about its edges. According to reporting from Axios, Anthropic has no plans to release “Model 2,” an internal model that appears to outperform Mythos, the company’s current top offering. Anthropic is not slowing development broadly — the work continues — but the company is prioritizing internal use over external release. The stated reasoning is risk management for advanced AI models.
There’s one interesting wrinkle in what Anthropic has chosen to talk about publicly. When the company discussed Mythos, it highlighted the model’s capabilities in one area in particular: identifying security vulnerabilities in software. That’s a telling choice. It’s a capability that cuts both ways — useful for defenders, dangerous in the wrong hands — and it hints at why the release calculus at Anthropic looks different than it does elsewhere.
Dario Amodei has also written that Anthropic continues to advocate for a judicious, evidence-based approach to these risks, even as he acknowledges the pendulum has swung and AI opportunity, not AI risk, is what’s driving the broader conversation in 2025–2026. In other words, Anthropic knows it’s making the unpopular argument right now, and it’s making it anyway.
What this means if you’re picking tools
Here’s my angle as a reviewer: the existence of an unreleased, stronger model changes how you should read the products you can actually buy.
First, the model you’re evaluating today is not the ceiling of what the company can do. That’s always technically true of every lab, but it’s rarely confirmed this directly. If you’re benchmarking Anthropic’s public offerings against competitors, remember you’re comparing their released tier, not their best tier. That’s neither good nor bad for your purchasing decision — you can only buy what’s for sale — but it matters for how you forecast the roadmap.
Second, a deliberate hold-back tells you something about how a vendor thinks. I’ve tested tools from companies that ship every capability the moment it compiles, and I’ve watched some of those launches go sideways. A vendor that keeps a stronger model internal because it’s worried about risk is, at minimum, a vendor that has a process for saying no to itself. Whether you find that reassuring or frustrating depends on what you’re building.
The case for frustration
If you’re a developer who wants maximum capability, this is annoying. Somewhere in Anthropic’s infrastructure sits a model that could presumably do your job better than the one you’re paying for, and the company has decided you can’t have it. Competitors who ship faster will happily take your subscription in the meantime. Caution has a real commercial cost, and Anthropic is choosing to pay it.
The case for reassurance
On the other side: if a lab publicly reports that risks are rising, and then acts consistently with that report by not releasing its most powerful system, that’s alignment between words and behavior. In my experience reviewing this industry, that alignment is rarer than it should be. Plenty of companies talk about safety in their blog posts and then ship whatever moves the revenue needle. Anthropic saying “risks are up, so this one stays inside” is at least a coherent position, whether or not you agree with it.
My honest read
I can’t review Model 2. Nobody outside Anthropic can. What I can review is the decision, and as decisions go, it’s a legible one: the company believes the risk profile of its strongest internal work doesn’t justify external release right now, and it’s willing to leave money on the table over it.
For toolkit buyers, the practical takeaway is unchanged in the short term — evaluate what’s shipping, test it against your workload, and ignore the ghost model you can’t access. But keep this episode in your notes. When a vendor’s internal capabilities and released capabilities diverge this openly, the gap itself becomes part of the story. I’ll be watching what eventually crosses that gap, and under what conditions. When it does, I’ll test it the same way I test everything else — skeptically, and hands-on.
🕒 Published: