Imagine a chef who perfects a new dish, plates it beautifully, photographs it for the menu — and then eats it alone in the back kitchen, telling customers the recipe isn’t safe for public consumption yet. That’s roughly where Anthropic has landed with its internal model reportedly called “Model 2.” The company says AI risks are rising, and it has no plans to release this model, which appears to be more powerful than its top-of-the-line Mythos. But it’s not slowing development. It’s just keeping the good stuff behind the counter.
As someone who reviews AI tools for a living, this puts me in a strange position. I can’t test what I can’t touch. And I’ve got some thoughts about what this move actually means — for Anthropic, for the industry, and for you, the person deciding which tools deserve your money.
What We Actually Know
Let’s stick to the facts, because there’s already enough vibes-based commentary out there. Anthropic has reported that AI risks are rising. The company is continuing to develop its technology, but it’s prioritizing internal use over external release for this stronger model. The stated reasoning is risk management — a cautious approach to handling more capable systems.
Meanwhile, Mythos exists as the company’s public flagship. Anthropic announced that model’s capabilities in one specific area: identifying security vulnerabilities in software. That’s a notable framing choice. Of all the things you could highlight about a frontier model, they picked the one that sits squarely at the intersection of “useful” and “potentially dangerous in the wrong hands.”
Caution or Competitive Strategy? Probably Both
Here’s my honest read. There are two ways to interpret keeping a stronger model internal, and they’re not mutually exclusive.
- The safety reading: Anthropic genuinely believes releasing a more capable model raises risks it can’t yet manage, so it’s holding back. This is consistent with the company’s long-standing public posture. Dario Amodei has continued to advocate for what he describes as a judicious, evidence-based approach to AI risks — even as he acknowledges the broader mood has shifted. By his own account, as of 2025–2026 the pendulum has swung, and AI opportunity, not AI risk, is what’s driving the conversation.
- The business reading: A more powerful internal model is also a competitive asset. If your best system helps you build your next products faster, you don’t need to sell access to it to extract value from it. Keeping it in-house isn’t just caution — it’s compounding advantage.
Neither reading makes Anthropic villainous. But the safety framing gets the headlines while the strategic benefit rides along quietly, and I think reviewers and buyers should keep both in view.
What This Means If You’re Actually Buying Tools
Practically speaking, nothing changes for you tomorrow. The model you can’t access is, by definition, not a tool you can evaluate. What you can evaluate is Mythos and the rest of the publicly available stack, and those should be judged on their merits — not on the halo effect of some stronger sibling locked in the basement.
This is where I’ll be blunt: “we have something better but you can’t have it” is a claim I can neither verify nor review. I’m not accusing anyone of exaggerating. I’m saying that from a buyer’s perspective, unreleased capability is worth exactly zero until it ships. Judge vendors on what they put in your hands.
Swimming Against the Current
What genuinely interests me is the timing. If the industry mood has swung toward opportunity and away from risk talk — as Amodei himself describes — then Anthropic publicly emphasizing rising risks and withholding a stronger model is a contrarian position. That takes some institutional spine, or at least a willingness to look unfashionable. In a market where every lab is racing to announce the next biggest thing, “we built it and we’re not releasing it” is an unusual sentence to say out loud.
Whether it ages as wisdom or as an overly conservative footnote depends on facts none of us have yet. For now, I’ll give Anthropic credit for consistency between its stated values and its release decisions, while keeping one skeptical eyebrow raised at how neatly caution and competitive advantage overlap here.
My job is reviewing what’s real and available. Model 2 is neither. Mythos is. That’s where my attention — and yours — should stay.
🕒 Published: