Open-weight models close the gap on reasoning benchmarks

Several open-weight releases now land within a few points of closed frontier models on multi-step reasoning, narrowing the capability divide.

Several open-weight model releases now sit within a few points of the leading closed systems on multi-step reasoning benchmarks. The gap that once separated the two groups has narrowed to a range many downstream tasks no longer care about.

Why it matters

When reasoning capability is available in weights anyone can download and run, the differentiator moves from raw model quality to fine-tuning, tooling, and serving cost. That changes where value accrues in the stack — toward the teams closest to the application and the data.

The numbers

On a common suite of multi-step reasoning evaluations, the top open-weight releases trail the best closed models by roughly three to five percentage points on average, down from double-digit gaps a year earlier. On several subtasks they are statistically indistinguishable.

Context

The convergence is driven less by any single breakthrough than by better training data curation, longer reasoning chains at inference time, and cheaper compute for the smaller labs that release open weights.

Background

Open-weight models have historically lagged on long-horizon, multi-step work. The current round of releases marks the first broad point at which that specific weakness is no longer reliable as a way to tell the two groups apart.