Anthropic’s Secret AI Researcher: Inside the 62.8% Automated Breakthrough

Anthropic has an internal AI that's automating the work of its own AI researchers with a SIXTY-TWO-POINT-EIGHT percent success rate. Last week, Admin, we talked about Anthropic signing its public work with invisible watermarks for authenticity. It turns out, while they were building that trust layer for the outside world, they were also building a private model that's starting to do the research FOR them. This isn't a product launch. This is a look inside the engine room. The model is codenamed, simply, Model 2. It was disclosed quietly in Anthropic’s August 2026 Risk Report. On paper, it sounds like a minor update. It scores just one and a half points higher than their best public model, Claude Mythos 5, on their internal capability index. But that number hides the real story.

The real story is what it can DO. On a benchmark designed specifically to simulate AI research automation, Model 2 hit that sixty-two-point-eight percent success rate. The next best model, Mythos Preview, was at fifty-four-point-eight. Their flagship, Mythos 5, was barely over fifty. That is a massive leap. This isn't about writing better poetry or summarizing emails faster. This is about an AI system becoming proficient at the very tasks required to build more advanced AI systems. Anthropic is already using it. Heavily. Model 2 is writing most of their production code. It’s running continuous, automated workflows. They've built a tool that is now accelerating its own development. This is the feedback loop everyone in the industry has been talking about, and Anthropic just put a number on it.

And yet, you can't use it. It's strictly for internal use, with no plans for a public release. They rate the safety risk as "low," but their caution speaks volumes. They’ve built a self-sharpening knife and they are keeping it firmly in their own workshop. So, while Anthropic builds its next-generation engine in a locked lab, what is OpenAI doing? They’re fortifying the castle they already own in public. Their top model, GPT-5.4, has been out since March. It's still the undisputed king of complex, multi-step tasks in the wild. The CEO of Zapier called it "the most persistent model to date," saying it finished jobs where every other model simply gave up. That’s their public-facing power. But their newest move, just yesterday, isn’t about power. It’s about trust.

OpenAI just rolled out a Zero Data Retention policy for enterprise customers. Here’s what that means: if you're a company using their API, they now guarantee your data is never stored after processing. It is NEVER used to train their future models. This is a direct shot at the single biggest fear holding back enterprise adoption. Every C-I-O worried about their proprietary data getting hoovered up into a future GPT-6 just had their main objection erased. This creates a fascinating split in the market. Anthropic is sprinting ahead on raw, internal R&D capability, effectively building an AI to accelerate its own progress. They are playing the long game, betting that the company with the fastest research loop will eventually win. OpenAI, on the other hand, is consolidating its current market lead.

They are making their state-of-the-art tools not just powerful, but also safe and trustworthy for the biggest, most risk-averse customers on the planet. They are building a moat of enterprise contracts and data security guarantees around their existing technology. One company is building the recipe for the next meal in secret. The other is turning their current kitchen into an impenetrable fortress that serves the entire world. The question for you is which front matters more. The quiet, compounding advantage of automated research, or the immediate, market-grabbing power of enterprise trust? Anthropic just showed its hand in the first game. OpenAI just raised the stakes in the second. The most important AI isn't always the one you can access. Sometimes, it's the one being used to build everything that comes next.