The Inference Report

AnthropicFable 5 holds the top position on SWE-rebench at 64.5 percent, unchanged from the prior round, with confidence intervals tight enough to rule out meaningful movement at the top tier. GrokGrok 4.5 remains at 63.8 percent and AnthropicOpus 5 at 63.4 percent, preserving the same three-model hierarchy. The interval widths across the top six entries (ranging from 0.54 to 1.83 percentage points) suggest that real separation exists between the first tier and everything below rank six, though the 1.41 percent margin around Fable 5 leaves room for natural variation on re-evaluation. On Artificial Analysis, the leaderboard shows broader movement within the 50 to 65 percent band where most capable models cluster, with Claude Opus 5 at 63.1 and Claude Fable 5 at 62.1, yet these two benchmarks measure different problem sets and evaluation protocols, making direct score comparison misleading. The SWE-rebench top ten remains stable in composition and order, suggesting that the coding task distribution and model capability gaps have not shifted materially. Lower-ranked models show more scatter, particularly between positions 220 and 430 on Artificial Analysis where single-point score differences become common and ranking precision degrades. The lack of disruption at the summit indicates that incremental model releases have not yet displaced the established leaders, and the consistency of the top six across both evaluation windows points to reliable separation rather than noise. Whether this stability reflects genuine plateaus in coding task performance or merely the lag between training and evaluation remains an open question.

Cole Brennan