On August 13, 2026, Google introduced Gemini 3.7 Flash — a model that DeepMind calls the "most intelligent workhorse model" for programming and AI agents. The release came just three weeks after Gemini 3.6 Flash and was the result of two factors: developer feedback and algorithmic innovations. The team took into account what the previous version was missing and significantly strengthened key scenarios — from code generation to document processing. At the same time, the starting price turned out to be half that of its predecessor.
What changed: key numbers
Progress is easiest to gauge through benchmarks. Across all key areas, the new model outperforms the previous version, and in some cases by a wide margin.
Code and engineering. FrontierCode 1.1 Main benchmark — 43.6% vs. 34.4%. This isn't just a percentage increase: the model solves tasks more accurately on the first pass, debugs code better, and finds errors more effectively. On DeepSWE v1.1, the result is 65.3% vs. 49.0% — generated code is more often production-ready, which is a key metric for real projects.
Web development. In the WebDev Arena ranking, the model scores Elo 1588 vs. 1538 for Gemini 3.6 Flash. The numbers don't grow as dramatically, but the quality changes: layouts and apps turn out more functional, and a reference — a screenshot, image, or design system description — is enough to follow the design accurately.
Documents and business processes. GDP.pdf shows 34.0% vs. 22.0% — nearly a 1.5x improvement when working with complex documents. AutomationBench — 30.4% vs. 17.0%: almost twice as good in real business scenarios. For routine automation, this is the difference between "trying it out" and "rolling it out."
Better at code and agent scenarios
What's behind the dry numbers? Developers get a model that behaves differently in practice. It adapts better to obstacles: if an agent hits an unexpected error, it doesn't crash but looks for a workaround. Gemini 3.7 Flash is more proactive in clarifying user intent, follows instructions more precisely, and puts more effort into multi-step planning and tool calls.
In practice, this means less manual oversight: no need to double-check every agent step or restart it after a failure. This is especially valuable in multi-agent scenarios. In demos, the model paired with Nano Banana generated a playable 3D game from a text prompt, worked with Gemini Omni and subagents to build interactive landing pages, and in a three-agent graph — trained robotics scenarios. What's impressive isn't so much the output itself as the degree of autonomy: the model maintains context and drives complex chains to completion.
Web and documents: where the gains are most visible
Document handling deserves special attention. The model turns a static PDF annual report into an interactive data story — this is no longer "extracting text" but full-fledged analysis and visualization. In web development, providing a reference is enough to get an accurate layout, with fewer prompts than before.
Price: half the cost with more capabilities
The most unexpected point is pricing. The introductory rate is $0.75 per million input tokens and $3.75 per million output tokens. For comparison, that's half the original cost of Gemini 3.6 Flash for the same volume. The promo price is valid until December 31, 2026, and will rise to $1.50 and $7.50 starting January 1, 2027 — still competitive for a model of this level.
For startups and indie developers, the price cut isn't just a nice bonus — it changes project economics: you can afford twice as many calls for the same budget or keep agents running significantly longer.
Personal AI agent gets a new engine
Starting from the announcement date, the personal agent Gemini Spark transitions to Gemini 3.7 Flash. It's a round-the-clock AI assistant launched at the I/O conference and available to Google AI Pro and Google AI Ultra subscribers in more than 160 countries. For them, the update means more efficient knowledge work, improved interaction with Google Workspace tools, and more accurate output in complex multi-step scenarios — such as consolidating files, drafting emails, or updating status documents.
Safety and availability
Expanded capabilities come with strengthened safeguards. The model received updated protections against abuse in chemical, biological, radiological, and nuclear threat areas, as well as against offensive cyber operations — while preserving beneficial use cases.
The new model has broad availability:
- For developers — via Google Antigravity, the Gemini API, Google AI Studio, and Android Studio.
- For businesses — via the Gemini Enterprise Agent Platform and the Gemini Enterprise app.
- For users — via Spark in the Gemini app for Google AI Pro and Ultra subscribers.
Summary
A rare case: a more powerful model simultaneously becomes more affordable. The twice-lower entry price changes the economics of agent solutions — for the same money, you can run twice as many parallel agents or run long call chains twice as often. And the benchmark gains show that cheaper here doesn't mean weaker: in code, documents, and web development, the new version confidently outperforms its predecessor. For teams automating routine work or building AI agents, this is a strong reason to test Gemini 3.7 Flash right now — especially since the promo price lasts through the end of the year.
