Gemini 3.7 Flash

What Gemini 3.7 Flash was

Gemini 3.7 Flash was a model Google released on 13 August 2026, three weeks after 3.6 Flash. Google described it as its most intelligent workhorse model yet for coding and agents^[1]. "Workhorse" is the label for the tier of a model family meant to be called in large volumes, as opposed to the largest and most expensive tier. That distinction sets what the release was about: not a push at the capability ceiling, but a step up in the model most developers actually call.

"Agents" here means models used to call tools and work through a task in several steps, rather than answer a single question in one pass.

How it compared with 3.6 Flash

Google published four comparisons against the model it replaced^[1]:

• FrontierCode 1.1 Main: 43.6%, against 34.4% for 3.6 Flash
• DeepSWE v1.1: 65.3%, against 49.0%
• WebDev Arena: 1588 Elo, against 1538 Elo
• AutomationBench: 30.4%, against 17.0%

The largest relative gain was on AutomationBench, which went from 17.0% to 30.4%, close to a doubling. DeepSWE rose by about 16 points and FrontierCode by about 9. WebDev Arena is scored in Elo rather than as a percentage, so the change there is 50 points on a rating scale. These evaluations cover the two things the release was pitched around: writing code and using tools.

Pricing

The introductory price, in effect through 31 December 2026, was $0.75 per 1 million input tokens and $3.75 per 1 million output tokens. From 1 January 2027 those rates rose to $1.50 and $7.50^[1]. Publishing both the current price and the date it changes is worth noting: anyone estimating the cost of a long-running workload had to account for a known date on which the bill would double. Per-token pricing is how these models are sold — you pay for the text going in and the text coming out, measured in tokens, the small units of text the model reads and writes.

Where it was available

The model was offered in Antigravity, the Gemini API, Android Studio and Gemini Enterprise, and to consumers through Gemini Spark for Google AI Pro and Ultra subscribers in more than 160 countries^[1]. Covering a coding tool, a raw API, an IDE, an enterprise product and a consumer subscription in the same announcement is the pattern for a model positioned as general infrastructure rather than as a single product.

The Gemini Spark update

Gemini Spark, the consumer surface, switched to 3.7 Flash at the same time, with improved tool use in Google Workspace apps^[1]. Google said the update improved accuracy and output quality on complex multi-skill workflows^[1] — tasks where the model has to chain several tools together rather than produce an answer in one step. That is exactly the kind of work the AutomationBench and DeepSWE figures describe, which is why the same release served both the developer and the consumer side.

A three-week lifetime

The release was superseded three weeks later by 3.8 Flash. gemini-3.7-flash was removed from AI Studio, and API requests naming it were auto-routed to gemini-3.8-flash^[2]. Google's Gemini API deprecations page listed the release date of gemini-3.7-flash as 13 August 2026^[2].

What it shows

Read on its own, 3.7 Flash is an incremental gain over its predecessor: better scores on coding and tool-use evaluations, a lower introductory price, wider distribution. Read as part of a sequence, it is also a record of release cadence. 3.6, 3.7 and 3.8 each arrived about three weeks apart, and the identifier developers typed into their code was retired roughly as quickly as it appeared. For anyone building on top of these models, the specific model name is not a stable foundation; the routing behaviour Google adopted, sending requests for a retired Flash model to its successor, is the mechanism by which that is handled^[2]. The archive value of an entry like this is that it pins down what was true on one date — the scores, the price and the endpoints — before the next release moved all three.