Anthropic's Model 2 AI, outperforming Claude Mythos 5, shown as a complex neural network diagram.

Editorial illustration for Anthropic's 'Model 2' Internally Outperforms Claude Mythos 5, Company Says

Anthropic's 'Model 2' Internally Outperforms Claude...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

August 20, 2026 • 3 min read

Anthropic is running an AI model in-house that beats every version of Claude the public can touch. The company disclosed this in its Risk Report from August 2026, tucked into a section on internal capability tracking rather than a product announcement. The model carries the codename "Model 2" and sits in what Anthropic calls the Mythos class, the same lineage as Claude Mythos 5.

There's no external release date, and Anthropic isn't saying when, or if, that changes. What makes this notable isn't secrecy for its own sake. Anthropic has built a habit of using Claude models internally before the public ever sees them, and this report gives a rare look at how it measures those unreleased systems against ones already shipped. The company tracks capability on an internal index called AECI, and it uses that scale to compare Model 2 against Mythos 5 in concrete terms.

The report also touches on how the model performs inside Anthropic's own operations, where Claude already handles a large share of production coding work. That internal reliance is part of why the company tests these models before anyone outside the building gets access.

Anthropic is running an unreleased AI model internally that outperforms every publicly available version of Claude. That's according to the company's Risk Report from August 2026.

Anthropic's most capable model, codenamed "Model 2," is for internal use only - THE DECODER

Why this matters

A 1.5-point AECI bump is the kind of number that would barely register as a version bullet point if Anthropic shipped it, yet the company is keeping Model 2 in-house while Claude Mythos 5 stays the public ceiling. That gap between what Anthropic runs internally and what it sells is worth tracking for anyone benchmarking against Claude or planning a product roadmap around its capabilities. If frontier labs are now sitting on models a full tier ahead of what customers can access, the published leaderboard stops being a real measure of the state of the art, it's a measure of what a lab is willing to release.

For developers and founders, that means discounting Anthropic's public capability claims slightly and asking what "internal only" actually screens for: safety evals, cost, or something else the Risk Report doesn't spell out. Researchers should watch whether Model 2 ever graduates to a public release, and how much the AECI gap widens before it does. Right now, the interesting story isn't Mythos 5, it's the model nobody outside Anthropic gets to touch.