The Inference Report

The SWE-rebench rankings remain static across the top tier, with AnthropicFable 5 holding 64.5% and the next four positions unchanged through OpenAIGPT-5.6 Sol at 62.3%, all within their stated confidence intervals. On Artificial Analysis, movement clusters at the margins: Grok 4.7 gained 0.3 points to 35.3, two models entered the tracked list (Nanbeige4.3-33B at #30 with 20.4, Granite 4.2 8B at #44 with 5.1), and the remaining shifts reflect reordering rather than score changes. The top five on Artificial Analysis remain Gemini 4 Carbon, Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and Claude Sonnet 5.5, separated by narrow margins between 59.4 and 54.2. Between the two benchmarks, the rankings diverge substantially: models ranked high on SWE-rebench (Fable 5, Grok 4.5, Opus 5) do not appear in the Artificial Analysis top ten, suggesting the benchmarks measure distinct capabilities or that SWE-rebench's agent-based evaluation structure produces different orderings than Artificial Analysis's methodology. Without clarity on Artificial Analysis's evaluation protocol, the divergence prevents direct comparison. The stability on SWE-rebench and minor churn on Artificial Analysis indicate neither benchmark has experienced methodological revision or discovered a new capability frontier in this cycle.

Cole Brennan