The Inference Report

The SWE-rebench rankings show no movement in the top tier: AnthropicFable 5 remains at 64.5 percent, with GrokGrok 4.5 and AnthropicOpus 5 holding positions two and three. The confidence intervals around these scores (ranging from 0.60 to 1.83 percentage points) suggest the top five models are clustered within a narrow band where day-to-day fluctuation is expected, and the current standings reflect measurement noise rather than meaningful separation. Across the Artificial Analysis benchmark, two entries have appeared since the previous cycle: Inkling 2 Small enters at rank 24 with a score of 30.6, and North Mini Code debuts at rank 58 with 3.4 points, pushing existing models down one position each. The remainder of the Artificial Analysis leaderboard has shifted uniformly downward in ranking, though scores themselves remain unchanged, indicating the new entries caused a cascade rather than any model improving or declining in absolute performance. Neither benchmark shows evidence of the field tightening or widening; the gaps between adjacent models persist as before. What has shifted is composition at the margins: the introduction of smaller or specialized models at the lower end of Artificial Analysis suggests broader participation in evaluation rather than advancement in the models already ranked.

Cole Brennan