Google has unveiled Gemini 3.7 Flash, a significant update to its Gemini 3.6 Flash model, designed to enhance reasoning capabilities and expand multimodal support. This latest iteration brings improvements to its core algorithm, enabling more sophisticated processing of text, images, audio, and video inputs within a 1 million token context window and supporting up to 64,000 tokens in output. Notably, the model also introduces customizable thinking configurations, allowing developers to tailor performance for specific use cases.
Coding and Agent Performance Boosts
One of the standout improvements in Gemini 3.7 Flash lies in its coding and agent capabilities. The model demonstrates a marked increase in performance on several benchmarks: achieving a 43.6% success rate on FrontierCode 1.1 Main, up from 34.4% in its predecessor, and scoring 65.3% on DeepSWE v1.1. Additionally, it reached a 1588 Elo rating on WebDev Arena, underscoring its growing proficiency in web development tasks. Document understanding and workflow automation also see notable gains, with GDP.pdf improving from 22.0% to 34.0% and AutomationBench rising from 17.0% to 30.4%.
Enterprise-Only Access and Pricing
While the model showcases impressive advancements, it is currently available exclusively through API and enterprise channels, with no open weights released to the public. The pricing is set at an introductory rate of $0.75 per 1 million input tokens and $3.75 for output tokens, valid through December 31, 2026. Despite its performance gains, the model still trails behind GPT-5.6 Terra in terminal and computer-use agent tasks, indicating room for further advancement in those areas.
Overall, Gemini 3.7 Flash positions Google at the forefront of multimodal AI development, offering developers and enterprises a powerful tool for complex reasoning and coding tasks, albeit with limited public access for now.
