Gemini 3.7 Flash Jumps Sharply on Coding Benchmarks, and the Launch Price Won't Last

What happened

Google released Gemini 3.7 Flash on August 13, just three weeks after Gemini 3.6 Flash. Flash is Google’s mid-tier Gemini model line: faster and cheaper than the top-end Pro models, built for the kind of high-volume, repeated calls a coding agent or automated workflow makes all day.

Google calls it “our most intelligent workhorse model yet for coding and agents,” and the benchmark gaps over 3.6 Flash are large enough to notice. On FrontierCode 1.1, a test of production-quality code output, Gemini 3.7 Flash scores 43.6% versus 34.4% for 3.6 Flash. On DeepSWE v1.1, which measures software engineering work that spans many steps instead of a single answer, it jumps to 65.3% from 49.0%. On WebDev Arena, a head-to-head ranking of web development output, its Elo score (a relative ranking score, higher means it wins more head-to-head comparisons) rises from 1538 to 1588. Document comprehension nearly doubles on the GDP.pdf benchmark, from 22.0% to 34.0%, and Google’s AutomationBench score for multi-step business workflows rises from 17.0% to 30.4%. These are Google’s own published numbers, not an independent test, but they’re a large enough jump across five different benchmarks to represent a real capability shift, not a rounding change.

The price is introductory, and it will double

Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. A token is roughly a word fragment, so a million tokens is a rough stand-in for word count. After that introductory window closes, the price doubles to $1.50 per million input tokens and $7.50 per million output tokens.

That’s worth knowing if you’re planning to build something on top of this model past the end of the year. The current price is a limited-time rate, not the model’s real long-term cost. If your usage volume is large enough that token cost shows up on a bill you actually watch, budget for the standard rate, not the launch rate.

Where you can use it

For developers, Gemini 3.7 Flash is available now in Google AI Studio (Google’s web console for testing prompts and calling the API, the interface that lets your own code send requests to the model directly), Android Studio, and Google Antigravity, Google’s coding tool that can plan and carry out multi-step tasks on its own instead of answering one prompt at a time. Enterprise customers get it through the Gemini Enterprise Agent Platform and Gemini Enterprise app. Consumers with an AI Pro or AI Ultra subscription can reach it through Gemini Spark, Google says, in more than 160 countries.

GitHub also added it to Copilot the same day, rolling out gradually to Copilot Pro, Pro+, Max, Business, and Enterprise plans. If you use Copilot, you’ll find it in the model picker in VS Code, Visual Studio, the Copilot CLI, the GitHub Copilot cloud agent, the GitHub Copilot app, JetBrains IDEs, Xcode, and Eclipse, once the rollout reaches your account. Copilot bills it at what GitHub calls “provider list pricing” under usage-based billing, meaning it draws from the same per-request allowance as other premium models rather than a flat subscription fee. If your Copilot plan is on a Business or Enterprise org, an admin has to turn on the “Gemini 3.7 Flash Preview” policy in Copilot settings before anyone in the organization can select it. If you don’t see the model yet and you’re on one of those plans, that policy toggle is the first thing to check, not a broken rollout.

Who should care

Anyone routing coding or agent workloads to a fast, cheap model tier should look at this release directly, especially if you were already using Gemini 3.6 Flash or a comparable model from another vendor for high-volume tasks. The benchmark gains suggest 3.7 Flash can now handle more of the multi-step work that used to require a slower, pricier top-tier model.

It matters less if you’re doing occasional, low-volume prompting where model choice barely touches your bill or your output quality either way.

What builders should do next

Pick one real coding or agent task you’ve already run on your current model, ideally something with a verifiable right answer, like a bug fix, a data extraction task, or a multi-step build. Run the identical prompt on Gemini 3.7 Flash through Google AI Studio or your Copilot model picker.

Compare four things: whether it got the task right on the first try, how many corrections you needed before the output was usable, how long the response took, and what it would cost you at $0.75/$3.75 per million tokens today versus $1.50/$7.50 after the introductory window ends. If it holds up on quality at the higher post-2026 price, it’s worth adopting now. If it only wins on the temporary launch price, treat that as a limited-time deal, not a permanent reason to switch.

End of article