Google Releases Gemini 3.7 Flash - Three Weeks After 3.6 Flash, at Half Price Through Year-End

Google released Gemini 3.7 Flash on August 13, 2026¹. The company positions it as “our most intelligent workhorse model yet for coding and agents”¹. It arrives just three weeks after the release of its predecessor, Gemini 3.6 Flash, and Google itself describes the release as coming “just three weeks after Gemini 3.6 Flash” and as “a direct result of developer feedback and algorithmic innovations”¹.

Pricing is set at an introductory rate available through the end of the year: $0.75 per million input tokens and $3.75 per million output tokens¹. Google describes this as “half the original 3.6 Flash cost per million tokens”¹. That price has an expiration date, however: a footnote states that introductory pricing expires on December 31, 2026, and that $1.50 per million input tokens and $7.50 per million output tokens will apply starting January 1, 2027¹.

The Gains Google Presented

Google presents benchmarks across four areas, but the comparison point in every case is the immediately preceding 3.6 Flash; no comparison figures against other companies’ models are given. All of the numbers below were presented by Google on its own blog and are not independently verified by a third party.

On coding, Google says 3.7 Flash shows gains over 3.6 Flash on tasks such as debugging and issue resolution, with higher first-pass code accuracy and improved generation of production-ready code¹. The figures are 43.6% on FrontierCode 1.1 Main (3.6 Flash: 34.4%) and 65.3% on DeepSWE v1.1 (49.0%)¹.

On web development, Google says the model produces functional layouts and feature-complete apps in fewer prompts, and that for UI generation it shows high design adherence whether the reference input is a screenshot, an image, or a full design system¹. On Arena.ai’s WebDev Arena, Google reports an Elo score of 1588 against 3.6 Flash’s 1538¹.

For knowledge-dense fields such as finance, law, and biosciences, Google cites 34.0% on the GDP.pdf benchmark, which tests processing of complex documents (3.6 Flash: 22.0%), and 30.4% on AutomationBench, which looks at completing real-world business workflows (17.0%)¹. Both evaluate handling of documents and multi-step business procedures rather than single-turn question answering.

On developer experience, Google says the model adapts better to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity, putting more effort into multi-step planning and tool calls¹. Google states that the resulting discipline means “less manual oversight and fewer retries across engineering workflows”¹.

The Introductory Price, and the Price It Returns To

The pricing is structured differently from a straightforward reduction. When this site covered the release of three Flash-tier models on July 21, 3.6 Flash was priced at $1.50 per million input tokens and $7.50 per million output tokens³. The figures in this announcement’s footnote for January 1, 2027 onward are the same numbers as that 3.6 Flash pricing¹³.

In other words, the $0.75 / $3.75 available this year is, as Google states explicitly, a time-limited introductory price rather than a permanent rate. In setups like agents, where a single task calls the model repeatedly, a difference in unit price flows directly into running costs. An annual cost estimated at this year’s rates will not hold from January 2027 onward. For anyone building a budget, both the introductory figure and the post-expiration figure need to sit side by side.

Competition on price continues regardless. As with OpenAI’s GPT-5.6 price cuts and Fast mode at the end of July, announcements from multiple companies have been staking out positions on cost per unit of performance rather than raw capability alone, and this release reads as part of that pattern. What distinguishes this one is that the reduction carries a stated expiration date.

Where It Is Available

Google lists availability in three groups¹. For developers, it points to Google Antigravity as a place to explore agent-first workflows, along with building in the Gemini API via Google AI Studio and Android Studio¹. For enterprises, access comes through the Gemini Enterprise Agent Platform and the Gemini Enterprise app¹. For individuals, it is available via Spark, the personal agent in the Gemini app, for Google AI Pro and Ultra subscribers in supported countries¹.

On Spark specifically, Google says it is available to Google AI Pro and Ultra subscribers in over 160 countries and began using Gemini 3.7 Flash the same day¹. That “over 160 countries” figure refers to Spark’s availability, not to the number of countries where the model itself or the API is offered. Google says the update makes Spark more efficient for knowledge work through improved tool use for Google Workspace apps, with better accuracy and output quality on complex, multi-skill workflows¹.

There is a route outside Google as well. In a changelog entry dated the same August 13, GitHub announced that Gemini 3.7 Flash can now be selected in GitHub Copilot². It covers Copilot Pro, Pro+, Max, Business, and Enterprise, billed on a usage basis at the provider’s list pricing². On Business and Enterprise, however, no one in the organization can select it until an administrator enables the relevant policy².

On safety, Google says it works continually to improve the coverage and robustness of its Frontier Safety safeguards, and that 3.7 Flash ships with updated safeguards against misuse in the domains of chemical, biological, radiological, and nuclear (CBRN) threats and cyber offense¹.

When a Generation Turns Over in Three Weeks

Three models — 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — shipped together on July 21³; three weeks later 3.7 Flash arrived, and it became selectable in GitHub Copilot the same day¹². At that cadence, an approach that pins one model and evaluates it exhaustively before adoption stops being practical. When lining up the pricing of the major tools, too, what matters is less the performance of any one model than how quickly a model can be swapped out and how easily a setup can be reworked when unit prices change.

Within what this announcement confirms, Google does not state the availability stage in the Gemini API — whether general availability or preview. GitHub includes “Preview” in its policy name², but it is safer to treat that separately from Google’s own positioning. Measuring the difference against 3.6 Flash on your own real tasks, before wiring the model into a production workflow, is the practical way to check the distance between published benchmarks and your own results.