Gemini 3.7 Flash: Benchmarks, API Pricing, and Availability

A workhorse model, three weeks apart

Google released Gemini 3.7 Flash on August 13, 2026, twenty-one days after Gemini 3.6 Flash. On FrontierCode 1.1 Main, a production code quality benchmark, the score rose from 34.4 to 43.6 percent. On DeepSWE v1.1, a long-horizon software engineering benchmark, it rose from 49.0 to 65.3 percent. Google also reported gains on WebDev Arena (Elo 1538 to 1588), GDP.pdf document processing (22.0 to 34.0 percent), and AutomationBench enterprise workflow automation (17.0 to 30.4 percent). [1][2]

Google's own reporting placed the new model's FrontierCode score narrowly ahead of Claude Sonnet 5, cited at 42.7 percent, and the-decoder separately reported that Google's testing showed the model outperforming GPT-5.6 Terra as well. Google attributed the pace of the jump to what it called algorithmic improvements rather than a change in model scale. [2][3]

Atlas interpretation: The three-week gap between releases is the more unusual fact than any single score. A workhorse tier moving ten points on a coding benchmark in that span reads less as a single model launch and more as evidence that Google is iterating its cheaper tier on a much faster loop than it ships anything else, which is itself a claim about where the model is made: mostly in training and evaluation pipelines rather than in a fresh pretraining run. [1][3]

Half price, aimed at coding and agents

Gemini 3.7 Flash launched at $0.75 per million input tokens and $3.75 per million output tokens, an introductory rate that holds through December 31, 2026, before standard pricing of $1.50 and $7.50 takes effect on January 1, 2027. Google described that introductory rate as half the cost per million tokens of Gemini 3.6 Flash, so the two models shared the same price through year end despite the capability gap between them. The model became available in Google AI Studio, Android Studio, the Gemini Enterprise Agent Platform, Gemini Spark, and Google Antigravity. [1][2]

Google positioned the release for coding and autonomous agent workflows, and said the model "better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity," which VentureBeat's reporting characterized as enabling more disciplined execution with fewer retries and less manual supervision. [1][2]

A fast tier while the flagship waits

VentureBeat reported that the release landed amid broader concern about Google's AI leadership, following delays to a Gemini 3.5 Pro update and the departure of researchers from the company's model teams. [2]

Atlas interpretation: That framing is the reason this Flash release is worth a page of its own rather than a line in the registry. A cheap, fast-moving tier shipping meaningful benchmark gains every three weeks is a visible, low-risk way for Google to keep showing progress while a slower-moving Pro-tier update has not shipped on the schedule some in the industry expected. The workhorse cadence does not resolve that question either way; it just gives Google something concrete to point to in the meantime. [2]