Gemini 3.7 Flash arrives half price, with the other half booked for January 1, 2027

Google launched Gemini 3.7 Flash on Thursday, three weeks after Gemini 3.6 Flash became available, positioning it as the company's workhorse for coding agents and automated business workflows ^[1]. It ships at half the original cost of Gemini 3.6 Flash, and that introductory rate ends on January 1, 2027, when input and output prices both double ^[2][3] - so any unit-economics spreadsheet built on the current number has a scheduled fracture in it.

The benchmark movement Google reports is real and uneven. DeepSWE v1.1 went from 49% to 65.3%, a 16.3 point gain and about 33% better in relative terms ^[4][1]. FrontierCode 1.1 Main went from 34.4% to 43.6%, up 9.2 points ^[5][2]. WebDev Arena moved from 1,538 to 1,588, a 50 point gain ^[6][3], which is the outlier in the set on the modest side. Outside coding, Google reports GDP.pdf rising from 22% to 34% and AutomationBench from 17% to 30.4% ^[7][8] - the AutomationBench figure is the largest relative jump Google published, roughly 79% ^[4]. These are vendor-reported numbers, and as The New Stack notes, benchmarks cannot show how reliably any of it holds inside a given company's own environment ^[9].

Now the part that survives contact with a finance review. Gemini 3.7 Flash currently costs $0.75 per million input tokens and $3.75 per million output tokens, a rate Google also applied to Gemini 3.6 Flash until December 31 ^[10][11]. On January 1, 2027 those become $1.50 and $7.50 ^[12]. Output is priced at five times input at both the discounted and the standard rate ^[5], which matters because the model's cost lever is a reasoning lever: thinking_level has low, medium and high settings, medium by default, and high exists so the model can spend more time reasoning and calling tools on hard coding, math and planning problems ^[13][14]. The setting that buys you the DeepSWE-style wins is the setting that inflates the expensive side of the invoice. The New Stack's framing is the right one: total token consumption across the full agent loop is what to measure, not the sticker rate ^[15]. A model that recovers from a bad first attempt can beat a cheaper-per-token competitor on the actual bill ^[16].

Migration is not free either. Teams coming from Gemini 3.5 Flash, Gemini 3 Flash Preview or Gemini 3.1 Pro have to remove temperature, top_p and top_k, replace the numeric thinking_budget with thinking_level, and drop candidate_count, according to Google's migration guide ^[17]. They also have to remove prefilled model turns, and Google recommends standardizing multi-turn interactions on the server-side previous_interaction_id; applications on the generateContent API need every FunctionResponse to include the matching call ID and function name ^[18]. That is a day of work for a small codebase and a sprint for a large one, and it is work you do twice if the next Flash arrives on a three-week cadence.

Google is not the only vendor buying adoption this way; OpenAI recently cut its own API prices amid rising competition, per the same report ^[19]. And this release is not the one that was expected: the model is Gemini 3.7 Flash, not Gemini 3.5 Pro ^[20].

Three things to watch. Whether the January 1 doubling actually lands or gets quietly extended, because a vendor that discounts to win agent workloads has an incentive to keep discounting. Whether Gemini 3.5 Pro shows up, and what that does to the Flash tier's positioning ^[20]. And, on your own side of the fence, whether you have measured tokens per completed task at each thinking_level before the price change rather than after.