جیمنی 3.7 فلیش

The next iteration in Google's Gemini 3 series of natively multimodal reasoning models, and what Google calls its most intelligent workhorse model yet for coding and agents. Algorithmic improvements to its core reasoning give substantial gains across software engineering, knowledge work and web development, with better debugging and higher first-pass accuracy on production-ready code than 3.6 Flash. It reads text, images, audio and video, and always reasons -- the thinking level is adjustable but cannot be switched off.

Flash is the middle rung of the Gemini ladder: more capable than the Flash-Lite models Google sells for throughput, cheaper and faster than the Pro line it sells for the hardest reasoning. 3.7 Flash is the top of that middle rung, and Google shipped it three weeks after 3.6 Flash rather than waiting for a generation boundary, describing it as the result of developer feedback and of algorithmic work it expects to carry into later models.

The gains Google publishes are concentrated in the places a workhorse model actually earns its keep. On FrontierCode 1.1 Main, which scores production code quality, the lab reports 43.6% against 34.4% for 3.6 Flash, and on the long-horizon software engineering set DeepSWE v1.1, 65.3% against 49.0%. On the head-to-head web development evaluation Google cites, it moves the Elo from 1538 to 1588 and generates feature-complete layouts in fewer prompts, including from a screenshot or a full design system as the reference. On agentic terminal work Google reports 85.8% against 78.0%.

Two of the largest jumps are in knowledge work rather than code. GDP.pdf, an evaluation of expert document comprehension, goes from 22.0% to 34.0%, and AutomationBench, which asks a model to complete real business workflows end to end, nearly doubles from 17.0% to 30.4%. Google also reports 97.0% on the eight-needle long context set at 128k against 91.8% for 3.6 Flash, which is the number to weigh against a million-token window.

Google spends as much of the post on behaviour as on scores. It says the model adapts to roadblocks, asks for clarification when a request is ambiguous, follows instructions more closely and puts more effort into multi-step planning and tool calls - which it frames as less manual oversight and fewer retries rather than as a benchmark result. The price is worth reading carefully too: Google launched it at half of what 3.6 Flash originally cost per million tokens, but that rate is introductory and runs to 31 December 2026, after which the lab has said input and output both double.

Where it stops is visible in Google's own table. On Terminal-bench 3.0, its broadest general-agent evaluation, 3.7 Flash scores 14.9%, and on OSWorld-2.0 computer use 47.9% - both far below the rows this model wins, and a reminder that autonomous desktop work is still early for every model in that table. The release also ships with tightened safeguards in the chemical, biological, radiological, nuclear and cyber-offence domains.