Via 9to5google.com
Three weeks after its predecessor couldn't produce a working file, Google's budget-tier AI model now zero-shots fully playable games, though it still trails smaller models in reasoning tasks.
Google DeepMind’s latest budget AI model can now build a playable video game from a single text prompt. Gemini 3.7 Flash, released on August 13, represents a remarkable turnaround from the Gemini 3.6 Flash launch just three weeks earlier, which struggled to produce even a functional output file.
The model is built for coding, web development, and complex multi-step workflows. Its headline trick, generating playable games from text descriptions, runs through Google’s Antigravity platform, which integrates with a real-time asset generation tool called Nano Banana to produce visual content on the fly.
Performance benchmarks tell a clear story of improvement. Coding efficiency jumped to 43.6%, up from 34.4% in the 3.6 Flash predecessor. That’s a roughly 27% relative improvement in under a month. Google also touts better debugging, production-ready code generation, design adherence, and more sophisticated multi-step planning capabilities.
The model is available across Gemini API, Google AI Studio, the Antigravity platform, and through the Gemini Spark agent for power users.
The pricing play #
Google is pricing Gemini 3.7 Flash aggressively. The introductory rate sits at $0.75 per million input tokens and $3.75 per million output tokens, a 50% cut from the previous release’s pricing. That promotional rate holds until December 31, 2026.
To put those numbers in context: a million tokens is roughly 750,000 words of text. At these prices, a developer could process the equivalent of about ten full-length novels worth of code for less than a dollar on the input side.
What it can’t do #
The caveat is significant and worth understanding clearly. Gemini 3.7 Flash is not a reasoning model. Google has positioned it explicitly as a fast, efficient coding tool rather than something that can think through complex logical problems.
The original coverage from Decrypt noted that a free 27-billion parameter model still writes better in certain tasks. That’s a telling admission about where Flash sits in the AI hierarchy. It’s optimized for speed and cost, not depth.
The three-week gap between 3.6 and 3.7 also raises questions about Google’s release cadence. Shipping a model that can’t produce working files, then following up with one that generates playable 3D games in under a month suggests either rapid internal iteration or a 3.6 launch that probably should have been delayed.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our