Google released Gemini 3.7 Flash on Thursday, three weeks after Gemini 3.6 Flash, and called it the company’s most intelligent workhorse model yet for coding and agents. The benchmark jumps are large for a three-week gap. On FrontierCode 1.1 Main it scored 43.6% against 34.4% for 3.6 Flash. On DeepSWE v1.1 it posted 65.3%. On AutomationBench — the multi-step agent test — it went from 17.0% to 30.4%, close to double.
Google says the model beats comparable releases from Anthropic and OpenAI across nine benchmarks, and that the biggest practical gains are in generating user interfaces: layouts produced by the model track reference images more closely than its predecessor’s.
The number that will actually change behaviour is the price. Input runs $0.75 per million tokens and output $3.75 through the end of 2026. On January 1, 2027 those become $1.50 and $7.50 — the introductory rate is exactly half the real one.
Cheap model, absent flagship
Gemini 3.5 Pro was announced at I/O in May 2026 for release the following month. It has not shipped. Google has issued no revised date and will not say whether it is still coming. That leaves the cheap tier two versions ahead of the flagship tier — an inversion of how model families are supposed to work, where the frontier model lands first and the distilled workhorse follows.
The distribution reflects the same priority. 3.7 Flash is live in Gemini Spark for Pro and Ultra subscribers, Google’s developer studio, Android Studio, the Gemini Enterprise Agent Platform, and GitHub Copilot for Pro, Business and Enterprise tiers. Every one of those is a place where a developer runs thousands of cheap calls, not one expensive one.
Our take: Build your cost model on $1.50 and $7.50, not $0.75 and $3.75. An introductory rate that doubles on a published date is a migration lever, not a discount — it is priced to make switching feel free during the months when switching costs are actually being incurred, and to reprice once the agent loops are wired in and nobody wants to rewrite them. If a workflow only clears its ROI bar at the introductory price, it does not clear the bar. Run the January numbers now, while you still have four months to move.
What to watch
- Whether 3.5 Pro ever ships. Three weeks of Flash cadence against a flagship three months overdue is a strategy statement, whether or not Google intends it as one.
- Agent benchmarks over code benchmarks. AutomationBench nearly doubling matters more than a single-shot coding score if you are running multi-step workflows, because that is where cost and failure compound.
- January 1 pricing behaviour. Whether Google holds the increase or quietly extends the intro rate will tell you how much of the volume was price-driven.
- The three-week clock. If 3.8 Flash lands in early September, the cadence is the product, and every evaluation you run has a shelf life measured in weeks.
The pattern to take from this week: the workhorse is where the competition now is, the flagship is late, and the sticker price on your invoice has an expiry date printed on it.