---
title: Anthropic Discloses an Unreleased Model It Calls Model 2
description: Anthropic's August 2026 Risk Report reveals an internal model, Model 2, that outperforms Claude Opus but has no plans for public release.
date: 2026-08-14T00:00:00.000Z
category: ai-news
tags: anthropic, claude, ai-safety
---

Anthropic disclosed the existence of an internal, unreleased model it calls Model 2 in its August
2026 Risk Report, according to [the report published on Anthropic's own
site](https://www.anthropic.com/aug-2026-risk-report), stating the model is somewhat more capable
than its current frontier model, Mythos 5, but that the company has no current plans to release it
externally.

## What Anthropic's report says

The report states that full pre-deployment evaluation of Model 2, including stealth side-task
testing, was not yet complete at the time of writing, so Anthropic has lower confidence in its own
assessment of the model's capabilities than it typically would for a model it plans to ship. Multiple
outlets covering the report, including [SiliconANGLE](https://siliconangle.com/2026/08/14/anthropic-details-unreleased-model-2-new-alignment-concerns-latest-ai-risk-report/)
and [Yahoo](https://tech.yahoo.com/ai/claude/articles/anthropic-sees-ai-risks-rising-191401564.html),
describe Model 2 as the first time Anthropic has publicly named an internal model it is choosing not
to release, rather than simply not mentioning it at all.

The same report raised Anthropic's own misalignment risk rating from "very low" to "low." Anthropic
says that change was driven by recent incident disclosures increasing overall uncertainty in its risk
estimate, not by anything specific to Model 2's capabilities, and that its underlying analysis likely
still supports a "very low" rating.

## Why it matters

Naming an unreleased model at all, rather than staying silent about it, reads as a deliberate
transparency move, and it likely reflects Anthropic trying to get ahead of the same disclosure
pressure that pushed OpenAI to publicly detail why it slowed development on its own Astra model this
month over cyber-risk concerns.

The bigger signal is what "not fully evaluated yet" implies about the pace of frontier capability
growth relative to safety-testing capacity. If Anthropic's own pre-deployment pipeline can't keep up
with what its research teams can already build internally, that gap is likely to recur at every major
lab racing toward the next capability tier, not just at Anthropic.

What to watch: whether Anthropic completes Model 2's evaluation and either releases a limited version
of it or discloses more about why it's holding back, and whether other labs follow with similar
unreleased-model disclosures of their own.

More [AI News](/ai-news/) coverage, or everything tagged [anthropic](/tag/anthropic/).