Google's answer to the frontier race is a cheaper Flash every three weeks

Google shipped Gemini 3.7 Flash on Thursday, August 13, a coding-focused model priced at half its predecessor and the company's second low-cost coding release inside three weeks ^[1][2]. The Gemini 3.5 Pro flagship it promised developers at I/O in May for a June launch still has not appeared ^[3].

The pitch is price. Through December 31, 2026, 3.7 Flash runs $0.75 per million input tokens and $3.75 per million output tokens, which Google says is half what 3.6 Flash cost at launch ^[4][5]. That implies a 3.6 Flash launch price of roughly $1.50 and $7.50 ^[6]. It is still not the cheapest option on the board: OpenAI prices the Luna version of its GPT-5.6 lineup at $0.20 and $1.20, per the same report, making Google's input tokens about 3.8 times and its output tokens about 3.1 times more expensive ^[7][8].

Google calls 3.7 Flash its strongest workhorse model for coding and agents, and attributes the gains to internal optimization work and developer feedback ^[9]. The benchmark set, credited to Senior Director Tulsee Doshi, is real movement rather than a rounding error: DeepSWE v1.1 from 49.0% to 65.3%, a gain of 16.3 points ^[10][11]; FrontierCode 1.1 Main from 34.4% to 43.6% ^[12]; WebDev Arena from 1,538 to 1,588, a 50-point move ^[13][14]. On document handling, the GDP.pdf benchmark went from 22.0% to 34.0% ^[15]. An AutomationBench workflow score is reported moving from 17.0% to 30.4%, though the source attributes that gain to 3.6 Flash rather than 3.7, which reads like a transcription slip ^[16].

None of that changes the standing. Gemini 3.7 Flash trails other major models on the Artificial Analysis Intelligence Index, where Anthropic's Claude Opus 5 leads ^[17], and the report's framing is that Google is behind Anthropic and OpenAI at the frontier while competing hard on price and cadence below it ^[18].

For operators, the consequence is a supply decision, not a capability one. Cheap agent tokens from a hyperscaler with distribution are useful, and 3.7 Flash is available through the Gemini API, AI Studio, Android Studio, Google Antigravity, and Gemini Enterprise ^[19]. But a three-week model cadence means your evaluation harness is stale roughly as fast as you can build it, and the current price is promotional, with a stated end date rather than a floor ^[2][4]. Google has not said whether newer versions of 3.5 Pro will arrive at all; it declined to tell Axios its plans, and reports suggest it may skip the model in favour of a future Gemini 4 Pro ^[20].

The consumer side shows how thin the rollout is. Individuals get 3.7 Flash only through the Gemini Spark agent on an AI Pro or Ultra subscription, while the standard Gemini chatbot still runs 3.6 Flash ^[21]. Google says the upgrade makes Spark better at multi-step tasks across Gmail, Docs, and other Workspace apps ^[22].

Three things to watch. Whether the $0.75 and $3.75 rates survive past the December 31, 2026 window or reset upward once workloads are migrated ^[4]. Whether 3.5 Pro ships, is renamed, or is quietly abandoned for Gemini 4 Pro ^[20]. And whether a 3.8 Flash lands in another three weeks, which would confirm the cadence is the product strategy rather than a gap-filler ^[2].