Anthropic is running its most capable model to date, internally codenamed "Model 2," entirely off the public grid: no API endpoint, no Claude.ai integration, no release date. According to The Decoder, the model is used exclusively by Anthropic's own team and has not been made available to customers or the public.

That's a notable gap. Every Anthropic flagship model to date has eventually landed in a customer-facing product, on a release cadence developers could roughly track. Model 2 breaks that pattern: it reportedly exists, it's described as the strongest system Anthropic currently operates, and there's no indication it's headed for a public release. It also raises a straightforward question for anyone tracking the frontier: how far ahead is Anthropic actually running, and how much of that gap will ever become visible?

For teams building on Claude, the useful question isn't what Model 2 can do — that detail isn't disclosed — but what its existence implies about the distance between what Anthropic has built and what it actually ships.

What's actually confirmed

The reporting is narrow, so it's worth being precise about what is and isn't known. Model 2 is:

Not known: parameter count, architecture, training data, intended purpose, or whether — and when — any version of it will reach the public. Anthropic hasn't commented beyond what surfaced in the report.

Why hold back your best model

Keeping an unreleased model running internally isn't unusual on its own — labs test ahead of what they ship, and Anthropic's own Responsible Scaling Policy requires internal evaluation before external deployment, which guarantees some lag between "built" and "released" by design.

This isn't unique to Anthropic — frontier labs typically run unreleased checkpoints internally, well before anything reaches paying customers. What makes this case notable is the vocabulary: Anthropic isn't describing Model 2 as a checkpoint under review, but as an active internal tool — a distinction that says more about intent than about safety process.

What stands out here, per the report, is the framing: Model 2 isn't described as a candidate sitting in a safety-review queue awaiting release — it's described as an internal tool the company is actively using. In our estimation, that reads more like a model built to accelerate Anthropic's own work — research, coding, training the next generation of Claude — than a finished product being held back solely pending a safety sign-off.

What it means for builders on Claude

Nothing changes in the API today — you still build against whatever Claude model is currently published. But the existence of Model 2 is a useful correction to how to read Anthropic's public roadmap:

None of this changes what you can call today. It's a reason to treat Anthropic's release calendar as a business decision layered on top of faster underlying progress, rather than a direct readout of the lab's current capability.

AiiN's takeaway

The notable part isn't that Anthropic runs an internal model ahead of its public releases — that's standard practice. It's that Model 2 is reportedly Anthropic's most powerful model, period, with no public counterpart at all right now. That's a wider gap between internal and external capability than Anthropic has previously made visible.

What's worth tracking is how much of Model 2 eventually surfaces — distilled, safety-filtered, or renamed — in the next Claude release. If that jump looks unusually large when it lands, this report is the reason why.