Google launches Gemini 3.7 Flash for coding and agents

Google has introduced Gemini 3.7 Flash, positioning it as the most capable model in its high-throughput Flash family for software development, AI agents, and demanding knowledge work. The release arrives only three weeks after Gemini 3.6 Flash and incorporates developer feedback alongside new algorithmic improvements that Google says may also influence future models. The central proposition is a combination of stronger performance and lower operating costs: during an introductory period running through the end of the year, the model costs $0.75 per million input tokens and $3.75 per million output tokens, which Google describes as half the original price of Gemini 3.6 Flash. Google is presenting Gemini 3.7 Flash as a production-oriented model that improves capability, reliability, and affordability at the same time.

The largest reported advances concern coding and software-engineering tasks. According to Google, Gemini 3.7 Flash is better at debugging, resolving issues, and producing correct code on its first attempt. It scored 43.6% on FrontierCode 1.1 Main, compared with 34.4% for Gemini 3.6 Flash, while its DeepSWE v1.1 result increased from 49.0% to 65.3%. Google interprets these results as evidence that the model can generate more production-ready software and reduce the number of corrections, retries, and interventions required from developers. The company also says the model responds to obstacles more effectively, asks for clarification when a request is ambiguous, follows instructions more faithfully, and devotes greater effort to multi-step planning and tool use.

Web development is another major focus. Gemini 3.7 Flash is designed to create more functional layouts and feature-complete applications with fewer prompts. When supplied with a screenshot, image, or complete design system as a reference, it is said to reproduce the intended interface with stronger visual adherence and closer functional parity. On Arena.ai’s WebDev Arena, its Elo score reached 1588, up from 1538 for Gemini 3.6 Flash. Google demonstrated applications ranging from a playable 3D game generated from a text prompt to interactive landing pages assembled through coordinated sub-agents. In these examples, Gemini 3.7 Flash acted as an orchestrator while other systems created visual assets, characters, textures, or parallax interface components.

The model also targets document-heavy professional work in finance, law, and biosciences. On GDP.pdf, a benchmark intended to test the processing of complex documents, Gemini 3.7 Flash scored 34.0%, versus 22.0% for its predecessor. Its AutomationBench score rose from 17.0% to 30.4%, suggesting improved performance on realistic business procedures that require several connected actions. One demonstration converted a static annual report into an interactive data story with live charts and aggregated findings. Another used the model’s multimodal understanding inside a three-agent loop to assist the training of a robotics model. The broader goal is to move beyond isolated answers toward systems that can interpret dense material, plan a workflow, call tools, and complete useful work with less supervision.

Google says the model’s more disciplined execution should make agent-based applications easier to operate at scale. Rather than merely producing text or code, Gemini 3.7 Flash is intended to recognize roadblocks, maintain intent across multiple stages, and select or invoke tools more reliably. This matters for production agents, where a small error at one step can disrupt an entire workflow. The company reports that early customers are seeing substantially better precision and outcomes than with Gemini 3.6 Flash while retaining a relatively low per-token cost.

Gemini Spark is among the first Google products to adopt the new model. Spark, introduced at Google I/O as a personal agent that can operate continuously under a user’s direction, is available to Google AI Pro and Ultra subscribers in more than 160 countries. With Gemini 3.7 Flash, Spark is expected to handle knowledge work more efficiently and use tools in Google Workspace applications more accurately. Google highlights practical tasks such as gathering files into one place, drafting emails, and updating project or status documents, including workflows that combine several different skills.

The release also includes revised safety measures. Google says it has strengthened protections intended to prevent misuse involving chemical, biological, radiological, and nuclear threats, as well as offensive cyber activity, while preserving legitimate uses under its bioresilience and cybersecurity programs. Additional technical and safety information is available in the model card. The safeguards reflect the increased autonomy of agent-focused models, whose ability to plan and use tools can amplify both useful and harmful actions.

Gemini 3.7 Flash is available to developers through the Gemini API in Google AI Studio and Android Studio, and it can be explored through agent-first workflows in Google Antigravity. Enterprise customers can access it through the Gemini Enterprise Agent Platform and the Gemini Enterprise application. Consumers can use it through Spark in the Gemini app if they have a Google AI Pro or Ultra subscription and live in a supported country. Together, these distribution channels show that Google intends the model to serve as a common foundation for developer tools, enterprise agents, and personal automation rather than as a narrowly focused coding release.