METR Analysis: Anthropic's New Model Clears 50% on Every Benchmark

Crypto Briefing reported on November 5 that a METR analysis of Anthropic's newest model found it scoring above 50% on all of the benchmarks included in the evaluation.

That number should be read as a floor, not a peak. Clearing half on every benchmark says the model has no category where it collapses; it does not say the model leads on any single one. The reporting available does not break out per-benchmark scores, task composition or sample sizes.

Third-party evaluation matters because vendors publish the numbers they choose to publish. Independent evaluators such as METR are watched closely because they run models against tasks the developers did not specifically design for, which makes results harder to optimize through data selection.

As a signal, uniform results above the halfway mark point to breadth. The gap between frontier models increasingly shows up on a small set of hard problems, while routine tasks get cleared by everyone; when a new model crosses the midline everywhere, it suggests the floor of baseline capability has risen again. For teams choosing a vendor, that typically shifts switching costs toward tooling and compliance rather than raw capability.

Three caveats are worth holding onto. Benchmark suites are finite, and the longer their items stay public the higher the contamination risk. A 50% line is an arbitrary threshold, and crossing it is not the same as reaching practical usefulness. And what exists here is secondary reporting, not the primary data.

What to watch: whether METR publishes the full write-up with task lists and per-benchmark results, whether Anthropic cites the analysis in its model card, and whether independent replications point the same direction. Until then, the defensible reading is that there is a broad-based result — not yet a definition of how strong the model is.