Testing Gemini 3.8 Flash on Coding, Vision, and Robotics

Bijan Bowen22:14
0 comments · 0 votesOpen discussionClose discussion
Sign in to join the discussion

    Video summary

    Bijan Bowen reviews Gemini 3.8 Flash as the third Flash release in six weeks, with performance improving while base pricing remains stable. The model accepts text, image, video, audio, and PDF inputs, offers a context window above one million tokens, and can use optional teamwork and deeper-reasoning modes.

    A browser operating-system test produced a feature-rich synthwave interface with games, mail, wallpapers, a code editor, and an animated assistant. Bijan Bowen liked the visual consistency and breadth, although the build also generated noisy notifications and several rough interaction details.

    The coding tests were mixed. A C++ skateboarding game improved substantially after visual feedback and a teamwork workflow, while a mechanical-watch page remained poorly composed and a subway game was playable but modest. A stronger reference image helped the model revise a game scene for more than an hour and reproduce much of its visual structure.

    The strongest result was a vision-guided robotic-arm task completed in 36 minutes. The system built a camera-feedback pipeline and pick-and-place script, though Bijan Bowen notes that an accidental movement may have made the final grasp easier. A Blender and Godot flight game completed quickly but looked simpler than requested and required a follow-up after a process stalled.

    Original YouTube thumbnailWatch on YouTube