Google reports 9.6% for Gemini 4 Carbon on SWE-sweep, a subtask-weighted reward it computed itself with a mini-swe-agent harness.
Tasks: Subtask-weighted reward, as Google reports it.
Setting: Google says its results use the Gemini API at the highest thinking setting unless noted; no numeric budget is stated.
Limits: Provider-reported: Google publishes the result for its own model; it has not been independently reproduced here. Google says the scores are as of October 2026 for the model it calls Gemini 4 Carbon; equivalence to the build now served on every surface is not established. Results under the same benchmark name can be shown together as reported scores, but unmatched settings and missing methods rule out a controlled ranking.
Google DeepMind · Google DeepMind · Google DeepMind
