METR's look at Anthropic's unreleased 'Model 4' shows outside evaluators now see the frontier before the public does

METR has identified internal Anthropic models that have not been released, according to a short video note. METR (Model Evaluation and Threat Research) is an independent group that tests AI systems for dangerous capabilities. The model it highlights is called 'Model 4'. The note says it was the highest scorer and passed 50% on every benchmark METR evaluated.

The source material does not say much more. It does not name the benchmarks or give exact scores. It does not compare Model 4 with Anthropic's public models, and it does not explain how METR got access. We won't fill those gaps. A score above 50% means little until we know which tests were used and how hard they are. Our own rule applies here: saturated benchmarks tell you nothing, and only hard, unguided tasks separate the models.

The score is not the main point. The main point is the gap between what labs run internally and what the public can use. Anthropic already leads our Zendoric Quality index with Fable 5. If unreleased systems are being tested above that public level, then the frontier as we chart it is a lagging indicator. The most capable models exist, and are evaluated, months before anyone outside can check them.

Our reading is cautiously positive. Outside evaluators testing unreleased models is the kind of evidence-based governance we keep asking for. Capability gets measured before deployment, not after an incident. The weak spot is the one we flagged with agent auditing: who certifies the result, and with what transparency. A single line saying a model 'beat 50% everywhere' is not accountability. It becomes accountability when the benchmarks, methods and risk findings are published in enough detail to scrutinise. Until then, treat 'Model 4' as a signal that the frontier is still moving. It is not proof of any particular capability.