Google's Gemini 3.7 Flash model can now turn a single prompt into a playable game. Three weeks ago, that same task was beyond it.
The model still lacks reasoning capabilities. And a free 27B parameter model writes better than it.
From prompt to playable
Zero-shot means the model receives one instruction and produces the game without any examples or prior training on that specific task. That's a big jump from where Gemini 3.7 Flash was three weeks ago, when it failed to complete the same kind of prompt.
There's no word on what kind of games it can generate, or how complex they are. The fact that it works at all marks a clear improvement.
Reasoning still out of reach
Despite the new trick, the model doesn't reason. That means it can't work through logic problems, multi-step calculations, or tasks that require careful deduction. It can produce output, but it can't think through the steps to get there.
That limitation separates it from models designed specifically for reasoning, and it's a gap Google hasn't addressed in this version.
Writing quality lags behind
Another check on the model's capabilities: its writing. A free 27B parameter model produces better text than Gemini 3.7 Flash. The comparison isn't close enough to call it a tie.
For a company pushing its AI into everyday tools, that's a notable weakness. The model can generate games, but it can't match the prose quality of a smaller, open-weights model.
Gemini 3.7 Flash can now make a game from a prompt. It still can't reason, and its writing trails a free alternative. Whether those gaps close in a future update is the open question.
