Three weeks. That's the lifespan Google gave Gemini 3.6 Flash this time around. On August 13 (Taipei time), the official blog posted the announcement for Gemini 3.7 Flash—barely three weeks after the previous generation's release, the new model was already live, officially billed as "the smartest flagship working model to date." But more than that description, what actually got the developer community talking was the pricing.
Gemini 3.7 Flash zeroes in on two things: writing code and running AI agent tasks. The benchmark numbers Google published show just how big the jump is—DeepSWE v1.1 scores leapt from 49.0% on 3.6 Flash to 65.3%, while FrontierCode 1.1 Main hit 43.6%. For developers, the more immediate feeling is that first-pass generated code is starting to hit production-ready accuracy. On the frontend side, just feed in a screenshot or a design system and the model can bring UI fidelity and consistency much closer to the original design.
The real killer move is the pricing. Starting now through the end of 2026—the "introductory pricing period"—Gemini 3.7 Flash's API costs are cut in half: just $0.75 per million input tokens and $3.75 for output. That half-price deal holds all the way until January 1, 2027, when it reverts to the standard rates of $1.50 and $7.50. In effect, Google just handed developers a year-and-a-half window to migrate their workflows into its ecosystem at a cost lower than the competition.
The new model has also been fed into Gemini Spark—the 24/7 personal AI agent service available to Google AI Pro and Ultra subscribers in over 160 countries worldwide, primarily handling cross-app workflows within Google Workspace like Gmail, Calendar, and Docs. On the safety front, Google has also pushed its Frontier Safety net a step further, updating protective mechanisms against risk categories including chemical, biological, radiological, and nuclear (CBRN) threats as well as cyberattacks. Developers can now access the new model directly through the Gemini API (Google AI Studio) and Android Studio.
