We’re happy to announce day-1 support of Gemini 3 Pro! Just boot up the OpenHands CLI and set gemini-3-pro-preview as the model: https://docs.openhands.dev/openhands/usage/run-openhands/cli-mode The model seems really good! With no tuning we got 70.4% on SWE-Bench verified, which means that with a bit of tuning we should be able to get a bit more.

why this is slightly lower than what Google claimed

The most likely suspect is that they use different file editing tools, there can be other small things. Also, maybe they ran experiments 100 times and then picked the best one? Not 100% sure 😅

@limited-student-10747 is it possible to add a col of the raw log of Gemini 3 pro? We want to take the raw log, measure total number of tokens and calculate the total cost of Gemini 3 Pro. (Column F)

Yes, but not till next week, we’re working on making the trajectories available