Google has introduced Gemini 3.7 Flash, calling it the most intelligent workhorse model yet for coding and agent-based tasks. The announcement was made on August 13, 2026, by Tulsee Doshi, Senior Director of Product Management, on behalf of the Gemini team.

Notably, the launch comes just three weeks after the release of Gemini 3.6 Flash. According to Google, this rapid follow-up reflects direct developer feedback along with new algorithmic improvements. The company says these advances will also carry forward into future models.

Gemini 3.7 Flash brings meaningful gains across several areas. These include software engineering, knowledge work, and web development workflows. Furthermore, Google is offering the new model at half the per-million-token cost of its predecessor, at least for a limited introductory period.

In terms of performance, Gemini 3.7 Flash shows clear improvements over 3.6 Flash on coding tasks such as debugging and issue resolution. The model also achieves higher first-pass code accuracy. On the FrontierCode 1.1 Main benchmark, it scored 43.6%, up from 34.4% for the previous version. Similarly, on DeepSWE v1.1, it reached 65.3%, compared to 49.0% before.

Moving to web development, Gemini 3.7 Flash generates more functional layouts and complete applications using fewer prompts. It also shows strong design adherence when working from a reference input, whether that’s a screenshot, an image, or a full design system. As a result, the model outperformed 3.6 Flash on Arena.ai’s WebDev Arena, posting an Elo score of 1588 against 1538.

In addition, the model shows improved reasoning for knowledge-heavy fields like finance, law, and biosciences. On the GDP.pdf benchmark, which tests a model’s ability to process complex documents, it scored 34.0%, compared to 22.0% previously. It also outperformed its predecessor on AutomationBench, a test of real-world business workflow automation, scoring 30.4% versus 17.0%.

Beyond the numbers, Google highlights several real-world demonstrations of the new model. These include generating a fully playable 3D game from a simple text prompt, using Gemini 3.7 Flash alongside Nano Banana to create characters, items, and textures in real time. Another demo shows the model orchestrating sub-agents with Gemini Omni to build interactive landing pages in a single shot. Elsewhere, the model was used to help train a robotics system through multimodal understanding in a three-agent graph loop. It was also shown converting static PDF reports into interactive data stories with live charts.

On the developer side, Google says Gemini 3.7 Flash offers a noticeably improved experience. It adapts better to roadblocks, clarifies intent when needed, and follows instructions more precisely. The model also puts more effort into multi-step planning and tool calls, which Google says results in less manual oversight and fewer retries during engineering work.

Pricing-wise, Gemini 3.7 Flash is available through the end of the year at an introductory rate of $0.75 per million input tokens and $3.75 per million output tokens. Combined with its performance gains, Google says this pricing allows developers and customers to scale production-ready agents more cost-effectively. Early feedback from customers, the company notes, has pointed to strong performance and precision at a relatively low cost.

Meanwhile, Gemini Spark, the personal AI agent available to Google AI Pro and Ultra subscribers in more than 160 countries, began using Gemini 3.7 Flash starting the same day. Spark was originally launched at I/O as a 24/7 personal agent capable of taking action on a user’s behalf. With this update, Google says Spark becomes more efficient at knowledge work, offering improved tool use across Google Workspace apps and better accuracy for complex, multi-skill workflows. As a result, Spark can now consolidate files, draft emails, and update status documents more efficiently.

On safety, Google states that Gemini 3.7 Flash ships with updated safeguards addressing misuse risks in chemical, biological, radiological, and nuclear domains, as well as cyber offense. These measures, the company says, align with its broader approach to bioresilience and its cyber safety program, while still supporting beneficial use cases. Additional details are available in the model card for 3.7 Flash.

Finally, access to the new model varies by user type. Developers can explore agent-first workflows in Google Antigravity or begin building through the Gemini API via Google AI Studio and Android Studio. Enterprises can access the model through the Gemini Enterprise Agent Platform and the Gemini Enterprise app. Individual users can reach it through Spark in the Gemini app, provided they hold a Google AI Pro or Ultra subscription in a supported country.

Overall, the rollout of Gemini 3.7 Flash marks another step in Google’s rapid iteration on its Flash model line, with the company positioning it as a faster, cheaper, and more capable option for coding and agentic tasks.

Related post

View all