The Inference Report

The SWE-rebench rankings show no movement at all in the top tier: AnthropicFable 5 remains at 64.5% ± 1.41%, holding first place with identical scores and confidence intervals across both snapshots. The top ten positions are frozen in place, from GrokGrok 4.5 at 63.8% down through CursorCursorAgent at 51.7%, with no changes in rank, score, or uncertainty bands. This stability at the coding task level contrasts sharply with the Artificial Analysis benchmark, where Solar Pro 4 dropped from position 78 at 26.6 points, creating a cascading reranking of nearly 400 entries below. The SWE-rebench data suggests either that no new models were evaluated on the code problem set since the previous snapshot, or that evaluation methodology remained sufficiently consistent to produce identical results with matching error margins. The coding benchmark's tighter confidence intervals (most under 1.5%) compared to the broader spread in general reasoning tasks may reflect more deterministic evaluation conditions, but the complete absence of movement warrants scrutiny: either the benchmark is genuinely stable and reproducible, or the underlying data collection has stalled. Without additional context on evaluation dates or methodology changes, the flatness of this leaderboard reads as a pause rather than a meaningful signal about model performance trajectories.

Cole Brennan