The Inference Report

The SWE-rebench ranking underwent wholesale turnover, with five new models entering the top tier and five prior leaders removed entirely. AnthropicFable 5 now leads at 64.5% ±1.41%, displacing OpenAI's gpt-5.5-2026-04-23-xhigh (which scored 62.7% ±0.91% previously and has been dropped), while GrokGrok 4.5 and AnthropicOpus 5 occupy the second and third positions at 63.8% and 63.4% respectively. The prior top performer, OpenAI's gpt-5.5-2026-04-23-xhigh, along with four other models designated as medium or xhigh variants, disappeared from the rankings entirely, suggesting either a methodology change, model retirement, or evaluation discontinuation rather than performance degradation. Among models that remained in the benchmark, JunieJunieAgent advanced from #2 to #6 despite a negligible score change (61.6% to 61.8%), while OpenAICodexAgent dropped sharply from #3 to #8 (60.4% down to 58.0%), and CursorCursorAgent fell from #9 to #10 (53.0% to 51.7%). Lower-ranked models show more volatile movement: MiniMax M3 climbed from #17 to #11 (45.6% to 47.2%), and MiMo V2.5 Pro rose from #19 to #12 (42.4% to 46.5%), while Qwen3.6-27B and Qwen3.6-35B-A3B both declined substantially, suggesting either model-specific improvements in the new cohort or changes in test conditions rather than uniform capability shifts. The confidence intervals remain narrow (typically ±0.5% to ±1.8%), indicating stable measurement precision, but the simultaneous departure of five prior leaders and arrival of five new ones raises a methodological question: whether this reflects genuine capability advances or a recalibrated evaluation protocol.

Cole Brennan