Gemini 3.7 Flash Ships 3 Weeks After 3.6, at Half Price

Google released Gemini 3.7 Flash on August 13, calling it "our most intelligent workhorse model yet for coding and agents." The notable part is not the version number but the interval: 3.6 Flash shipped three weeks earlier. Google credits the compressed cadence to developer feedback and what it describes as algorithmic innovations rather than a larger training run.

The reported gains are concentrated in code and agentic work. On FrontierCode 1.1, Google puts 3.7 Flash at 43.6% against 34.4% for 3.6 Flash. On DeepSWE v1.1, the long-horizon software-engineering benchmark, the score moves from 49.0% to 65.3%. WebDev Arena Elo rises from 1538 to 1588, the GDP.pdf document-reasoning set goes from 22.0% to 34.0%, and AutomationBench — the roughest measure of whether a model can drive a multi-step workflow unattended — nearly doubles from 17.0% to 30.4%. Google frames the difference in behavioral terms, saying the model "thinks more diligently, putting in more effort into multi-step planning and tool calls," with more disciplined execution that needs less manual oversight.

Pricing is where the release bites hardest. The introductory rate is $0.75 per million input tokens and $3.75 per million output tokens, held through December 31, 2026, after which it reverts to $1.50 and $7.50 — meaning the launch price is half the standard rate, and roughly half what 3.6 Flash cost per million tokens. That puts a model Google positions near the frontier on coding into the price bracket enterprises had been reserving for cheap classification and summarization work. InfoWorld read the move as a deliberate divergence in enterprise AI economics, with the cheap tier iterating fast while the Pro cadence slows.

Availability is broad from day one: Google AI Studio and Android Studio for developers, the Gemini Enterprise Agent Platform and Gemini Enterprise app for companies, and Gemini Spark for Google AI Pro and Ultra subscribers across more than 160 countries.

Cross-vendor comparisons published alongside the launch put the picture in context, and it is not a clean sweep. On the FrontierCode 1.1 main track, 3.7 Flash leads at 43.6% against 42.7% for Claude Sonnet 5 and 41.3% for GPT-5.6 Terra — a margin thin enough to sit inside run-to-run noise. On DeepSWE v1.1, GPT-5.6 Terra still leads at 69.6%, with 3.7 Flash at 65.3% and Meta's Muse Spark 1.2 at 59.3%. Every one of these numbers comes from vendors with an interest in showing capability climbing, and none of these suites carry independent academic audit, so the honest reading is directional rather than decisive.

What the release does settle is the shape of the competition. A three-week turnaround on a mid-tier model, shipped with a price cut instead of a premium, suggests Google now treats Flash as the volume battleground rather than a stripped-down afterthought to Pro. Whether the January 1 reversion to full pricing holds, or gets quietly extended once rivals answer, is the thing worth watching.