Google launches Gemini 3.7 Flash for AI agents

Google has launched Gemini 3.7 Flash, an AI model designed for coding, tool use and completing multi-step tasks. It arrives just three weeks after Gemini 3.6 Flash and, through the end of the year, costs half as much per million tokens processed.

The new version targets developers and companies building agents: systems that can plan a task, retrieve information, call other tools and take actions with limited supervision.

More accurate coding and app creation

Google says 3.7 Flash improves on 3.6 Flash in debugging, error resolution and production-ready code generation. In its tests, it achieved these results:

• 43.6% versus 34.4% on FrontierCode 1.1 Main.
• 65.3% versus 49.0% on DeepSWE v1.1.
• An Elo score of 1588 versus 1538 on WebDev Arena, an evaluation of website and web app creation.

In practice, this means it can deliver a functional first version with fewer attempts. It can also reproduce an interface from a screenshot, image or design system while more closely preserving the specified structure and appearance.

It does not eliminate the need to review the code. But it can reduce the time you spend fixing basic errors or repeating instructions so the model completes a task.

Better performance on documents and workflows

The model also targets information-heavy tasks, such as reviewing financial, legal or scientific documents. In the GDP.pdf test, designed to measure understanding of complex documents, it achieved 34% versus 22% for Gemini 3.6 Flash.

On AutomationBench, an evaluation of real business processes, it scored 30.4% versus 17%. Google presents these results as a sign that 3.7 Flash can complete more steps in a workflow without constant intervention.

The model also aims to clarify ambiguous instructions more effectively, adapt when it encounters an obstacle and plan before making tool calls. For an agent, that difference matters: fewer retries and less manual supervision can reduce the cost of running it for hours.

Pricing and availability

Through the end of the year, the introductory price is $0.75 per million input tokens and $3.75 per million output tokens. Tokens are the units the model uses to process text, instructions and responses.

Developers can try it through the Gemini API, Google AI Studio, Android Studio and Google Antigravity. Companies can access it through Gemini Enterprise Agent Platform and the Gemini Enterprise app.

Gemini Spark also gets the model

Gemini Spark, Google's personal agent that can take actions at the user's direction, starts using Gemini 3.7 Flash today. It is available to Google AI Pro and Ultra subscribers in more than 160 countries, where supported.

With the new model, Spark can handle tasks that combine several Google Workspace capabilities more accurately, such as gathering files, drafting emails and updating status documents.

Google also says it has updated safeguards against misuse in chemical, biological, radiological, nuclear and cyberattack-related areas. The model is being released with a safety card detailing these measures.

The important news is not just that Gemini is more capable. Google is trying to make agents reliable and inexpensive enough to use in production. The question to watch now is whether the improvements seen in testing hold up in real-world tasks, where documents are messy, instructions change and mistakes have consequences.