Google Unveils Gemini 3.7 Flash: A Leap Forward in Low-Cost Agentic Models

Google has launched Gemini 3.7 Flash, an updated version of its low-cost agentic model. This new iteration comes just three weeks after the release of Gemini 3.6 Flash and focuses on coding, agentic tasks, and complex knowledge-work scenarios.

Gemini 3.7 Flash demonstrates significant performance improvements over 3.6 Flash in software engineering, web development, and knowledge-intensive tasks. In terms of coding capabilities, it better handles debugging, issue fixing, and production-grade code generation tasks.

The new model has shown improved results across multiple tests, including the FrontierCode 1.1 Main benchmark, where its score was 43.6%, higher than 3.6 Flash's 34.4%. In the DeepSWE v1.1 long-horizon software engineering test, it scored 65.3%, also higher than the previous version's 49.0%.

Gemini 3.7 Flash can generate more complete applications and interactive pages with fewer prompts and replicate UI styles based on screenshots, images, or full design systems. Its Elo rating in the Arena.ai WebDev Arena is 1588, higher than Gemini 3.6 Flash's 1538.

The new model also shows improvements over the previous generation across multiple tests, including coding arenas, enterprise workflow automation, and complex document understanding. In scenarios requiring the processing of complex materials, such as finance, law, and life sciences, Google said the new model's reasoning accuracy has been further enhanced.