Gemini 3.6 Flash

A high-efficiency Gemini providing sustained frontier-level intelligence for real-world tasks at higher speed and lower cost. Google built it for the agentic era, and it excels at code generation, agentic execution and spatial reasoning while using about 17 percent fewer output tokens than the previous Flash generation at better quality. Same four input modalities and million-token context as 3.7 Flash.

Google released 3.6 Flash as the workhorse of a three-model announcement: the Flash tier for general production work, a cheaper 3.5 Flash-Lite underneath it for throughput, and a restricted 3.5 Flash Cyber for vulnerability work that Google keeps to governments and trusted partners. Its stated design goal was not a higher peak score but a lower cost per completed agentic task, which is a different thing from a lower price per token.

That is why the efficiency claim leads the post. Google reports the model spending fewer output tokens than 3.5 Flash for better answers, and taking fewer reasoning steps and tool calls to finish a multi-step workflow, while charging less per token than the model it replaced. A cheaper token that is spent twice as often is not cheaper, and the lab measures the difference at the task level.

On capability the lab reports higher precision with fewer unwanted code edits and fewer execution loops on DeepSWE, at 49% against 37% for 3.5 Flash; a large jump in machine learning research work on MLE Bench, at 63.9% against 49.7%; better computer use on OSWorld-Verified, at 83.0% against 78.4%; and better knowledge work on the GDPval-AA v2 Elo, at 1421 against 1349. Computer use also became a built-in client-side tool with this release rather than something the caller has to assemble.

The multimodal side is where Google points customers with document workloads: parsing documents, reading charts and data, and drafting reports off them. It ships with strengthened safeguards in the chemical, biological, radiological, nuclear and cyber-offence domains, which Google says make it substantially harder to jailbreak while being trained to refuse less on legitimate requests.

The honest caveat is the calendar. Gemini 3.7 Flash arrived three weeks later at an introductory price Google set at half of 3.6 Flash's original rate, and the lab's own comparison puts it ahead of 3.6 Flash on every coding, document and workflow row it publishes. 3.6 Flash remains a reasonable choice where a workload has already been tuned against it, not where a new one is being started.