Google's Gemini 3.7 Flash undercuts its own predecessor by half on price
The workhorse model posts 65.3% on DeepSWE v1.1 against 49.0% for 3.6 Flash, and runs at $0.75 per million input tokens until the end of the year.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Introductory pricing is $0.75 per million input tokens and $3.75 per million output, half what 3.6 Flash cost.
- Google reports 43.6% on FrontierCode 1.1 against 34.4% for the previous model, and 65.3% on DeepSWE v1.1 against 49.0%.
- Regular pricing of $1.50 and $7.50 per million tokens takes effect on 1 January 2027.
Google released Gemini 3.7 Flash on 13 August, describing it as its most intelligent workhorse model yet for coding and agents. The release comes three weeks after 3.6 Flash, a cadence that now looks less like a product cycle and more like a patch schedule.
The pricing is the headline. Introductory rates run at $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026 — half the per-million cost of 3.6 Flash by Google's own reckoning. Regular pricing of $1.50 and $7.50 begins on 1 January 2027, so the discount is a nine-week window, not a permanent repricing.
The benchmark gains Google reports are large enough to be worth stating precisely: 43.6% against 34.4% on FrontierCode 1.1 for production code quality, 65.3% against 49.0% on DeepSWE v1.1 for long-horizon engineering, an Elo of 1588 against 1538 on WebDev Arena, 34.0% against 22.0% on a document-processing benchmark, and 30.4% against 17.0% on AutomationBench. All are Google's own figures against its own previous model.
That last point is the caveat. A vendor comparing a new model to its predecessor controls both sides of the comparison, and none of these numbers place 3.7 Flash against a competitor. The useful reading is the direction and size of the jump on agentic and long-horizon tasks, which is where the previous generation was weakest.
Halving the price of a competent coding model for nine weeks is a customer-acquisition move aimed squarely at whoever is currently paying more for agentic workloads. The more durable signal is the AutomationBench jump from 17.0% to 30.4%: business-workflow automation is the use case labs have been promising for two years and mostly failing to deliver, and a near-doubling on that axis in three weeks is the number to watch.