Gemini 3.7 Flash arrives three weeks after 3.6, at half the token price

Google released Gemini 3.7 Flash, which it calls its most intelligent workhorse model yet for coding and agents ^[1]. The release landed three weeks after Gemini 3.6 Flash, at an introductory price Google says is half the per-token cost of 3.6 ^[2][3]. For teams that pin a specific model version into an agent stack, a three-week cadence at a falling price turns re-benchmarking and re-pricing from a one-time integration into a recurring operations task.

The published price is $0.75 per million input tokens and $3.75 per million output tokens, held through the end of the year ^[4]. If that is half of 3.6 Flash as Google states, the prior model ran at roughly $1.50 input and $7.50 output per million tokens ^[5]. The "introductory" and end-of-year framing means the number is not a fixed input to a budget; a team that sized its unit economics on the launch price is exposed to a change once the introductory window closes ^[4].

The benchmark deltas Google cites are not marginal. On FrontierCode 1.1 Main the company reports 43.6 percent versus 34.4 percent for 3.6 ^[6], and on DeepSWE v1.1 65.3 percent versus 49.0 percent ^[7]. WebDev Arena Elo moves from 1538 to 1588 ^[8]. On document-processing and workflow evals the gaps are wider: GDP.pdf at 34.0 versus 22.0 percent, about a 55 percent relative gain ^[9][16], and AutomationBench at 30.4 versus 17.0 percent, close to an 80 percent relative jump ^[10][17]. Google attributes the improvements to developer feedback and algorithmic changes ^[11].

Bigger jumps are the operational problem, not the reassurance. A model that scores substantially higher on issue resolution and business-workflow tasks will behave differently on the prompts and tool chains a team has already tuned, so the safe assumption is that pinned behaviour changes with every point release. Google's own description supports this: it says 3.7 Flash adapts to roadblocks differently, clarifies intent, and follows instructions with greater fidelity, which is a change in agent behaviour, not just accuracy ^[15].

There is a tell about how fast Google intends to rotate the tier. Gemini Spark, the 24/7 personal agent launched at I/O and available to AI Pro and Ultra subscribers in over 160 countries, switched to 3.7 Flash the same day ^[12][13]. The model also ships with updated safeguards against misuse in chemical, biological, radiological, nuclear, and cyber-offense domains ^[14].

What to watch: whether the three-week gap becomes the standing cadence or was a one-off, and what happens to the $0.75/$3.75 price when the introductory period ends ^[2][4]. Teams running agents in production should assume both the price and the model's behaviour are variables to re-measure on a monthly cycle, and budget engineering time for regression testing rather than treating a version pin as stable.