The Inference Report

On the SWE-rebench, the top tier remains locked in place: OpenAI's gpt-5.5-xhigh holds 62.7% plus or minus 0.91%, followed by JunieAgent at 61.6% and OpenAI CodexAgent at 60.4%, with no movement among the leading six entries. Below that tier, Z.ai's GLM-5.2 enters at position 12 with 51.1% plus or minus 1.13%, displacing its predecessor GLM-5.1 to 13th place, while DeepSeek-V4 Pro and MiMo-V2.5-Pro appear as new entries at 18 and 19 respectively, and Qwen models now occupy positions 22 and 23 in their first SWE-rebench appearances. The Artificial Analysis benchmark shows broader volatility: Claude Fable 5 leads at 59.9, a model not previously ranked in the earlier snapshot, while GPT-5.1 dropped from 38.9 to 36.9 (position 44), and Command A+ fell from 29.3 to 22.5 (position 111), the largest documented decline. gpt-oss-20b, which had held position 171 at 14.9, has been removed from the rankings entirely. The SWE-rebench data carries tighter confidence intervals than Artificial Analysis, suggesting more controlled evaluation conditions, though both benchmarks show the frontier remains dominated by OpenAI and Anthropic systems when measured on code completion tasks, with newer Chinese models (Qwen, GLM variants) gaining ground in the mid-tier rather than displacing leaders.

Cole Brennan