November 2025 Power Ranking

The November Power Ranking is here, with new models from all the big labs!

Brokk's Power Ranking consists of real-world coding tasks in large Java codebases, containing only tasks recent enough that they have not yet been trained against. Compared to SWE-Bench or LiveCodeBench, this gives a much more accurate picture of models' coding ability when faced with a problem they haven't seen before.

Commentary

Haiku stands alone in the S Tier. This is the most subjective call in the Power Ranking, but we think Haiku deserves to be recognized for its blisteringly fast responses while being smart enough to solve most problems correctly. (And when it can't solve your problem, you get the feedback quickly enough that you can switch to a smarter model without first losing your entire morning.) I wrote in October that I was making it my primary coding model, and it still holds that place.

GPT 5.1 and GPT 5-Mini take places in the A Tier for the straightforward reason that these are the only models on the Pareto front for both speed and cost. GPT 5.1 (with priority processing) is my preferred planning model. Even with priority processing, it's slower than Sonnet, but it's enough smarter that it's worth it; with Sonnet I have to redo enough work that it doesn't save me time overall.

Sonnet 4.5 and Grok 4.1 Fast make the B tier as the other models on the Pareto front, for speed and cost respectively. GLM 4.6 also gets a spot in B as the strongest open-weights model.

Both of the Gemini models are relatively expensive and slow for their level of intelligence, so they both hit the C Tier with Opus 4.1. We expect the full Gemini 3 Pro release to do better after resolving the preview's teething pains.

Finally, K2-Thinking and M2 land in the D Tier with thoroughly lackluster results.

Check out the full Power Ranking results here.