- In its August 2026 risk report, Anthropic disclosed two unreleased internal models, Model 1 and Model 2, and said Model 2 is more capable than Claude Mythos 5, its current public flagship.
- Anthropic raised its own misalignment risk rating for high-stakes autonomy from very low to low, stating it did so to reflect increased uncertainty, not confirmed danger.
- On the SHADE-Arena sabotage test, Mythos 5 recorded under a 1% stealth success rate with extended thinking, down from the 6 to 10% range of the earlier Mythos Preview.
Anthropic disclosed a model more capable than its public flagship
Anthropic's latest risk report, a redacted document running to roughly 186 pages, revealed that the company is running two unreleased successors to Claude Mythos 5 internally. It calls them Model 1 and Model 2. Model 1 is described as broadly similar to earlier releases, while Model 2 is a noticeable improvement on Mythos 5 for many tasks relevant to internal use, though the report is careful to say it is not the kind of leap seen between Opus 4.6 and Mythos Preview.
Model 2 is not a lab curiosity. Anthropic says staff use it heavily for software development, generating training data, and automating engineering work, the tasks that most directly speed up the company's own research. What the public can buy today is a step behind what the company builds with internally.
The more striking disclosure is about risk. Anthropic raised its assessed misalignment risk for high-stakes autonomy from very low to low, and it was explicit that the change was driven by uncertainty rather than evidence of harm.
We believe that the arguments presented below likely still support a designation of very low risk, but we are raising our assessed risk to low to reflect increased overall uncertainty.Anthropic, August 2026 Risk Report
| Unreleased internal models disclosed | 2 (Model 1 and Model 2) |
| Misalignment risk for high-stakes autonomy | Raised from very low to low |
| Mythos 5 stealth success on SHADE-Arena | Under 1% with extended thinking |
| Length of the redacted report | ~186 pages |
Why a self-imposed risk raise matters more than the model
Companies almost never volunteer bad news about their own products. A frontier lab publishing a report that raises its own risk rating, while sitting on a more capable model it has chosen not to ship, is a governance signal worth more than the specifications of Model 2.
The reason for the raise is the part to sit with. Anthropic did not say its models became more dangerous. It said it is less confident that its own evaluations keep pace with how fast the models are improving, and flagged that it is rewriting its threat models in light of recent incidents. That is a lab admitting its measuring instruments may be lagging the thing they are meant to measure.
The most important line in the report is not about Model 2. It is Anthropic conceding that it raised its own risk estimate because it is no longer sure its tests still hold. When the companies building the most capable systems say their safety measurements are struggling to keep up, the confidence gap, not the capability gap, is the thing to watch.
Santage is committed to independent, transparent journalism. This article is produced in accordance with Santage's Editorial Standards and aims to provide accurate and timely information. Readers are encouraged to verify information independently.