Google has introduced Gemini 3.6 Flash and 3.5 Flash-Lite as production workhorses for agents that must run quickly and repeatedly. A restricted Gemini 3.5 Flash Cyber variant adds a third path for vulnerability remediation.
The releases address an economic reality often hidden behind model benchmarks: an agent may invoke a model many times for one task, so unnecessary reasoning tokens add cost and latency at every step.
Gemini 3.6 Flash aims to do more with fewer tokens
Google says Gemini 3.6 Flash produces 17% fewer output tokens than 3.5 Flash on measurements from the Artificial Analysis Index. On individual synthetic tests, including DataCurve DeepSWE, the reduction reportedly reaches 65%.
Pricing is $1.50 per million input tokens and $7.50 per million output tokens. The model targets coding and multimodal reasoning where agents must plan, call tools and process varied data without the cost profile of a flagship model.
Google reports a DeepSWE success rate of 49%, up from 37%, and an MLE Bench improvement from 49.7% to 63.9%. Its GDPval-AA v2 score, intended to reflect knowledge work, rises from 1,349 to 1,421. These benchmarks suggest better capability, but teams should reproduce their own tool-use and document tasks.
Figma is using 3.6 Flash in prototyping, while Harvey and Hebbia apply it to financial filings, charts and draft reports. Google has also embedded a client-side computer-use tool in the Gemini API and enterprise platform, reducing the custom middleware required to operate software interfaces.
The company reports an OSWorld-Verified score of 83%, compared with 78.4% previously. It also claims stronger safeguards against chemical, biological, radiological and nuclear misuse without increasing benign refusals. Production buyers need detailed evaluation for their own operating environments rather than treating aggregate scores as guarantees.
Flash-Lite handles volume while Cyber stays restricted
Gemini 3.5 Flash-Lite is designed for background work such as document processing and agentic search. Artificial Analysis measured 350 output tokens per second, which Google describes as the fastest performance in the 3.5 family.
At $0.30 per million input tokens and $2.50 per million output tokens, it can handle routine subagent calls while a more capable model receives difficult steps. Its long-context result on Google’s GDM-MRCR v2 test rose from 60.1% to 72.2%, and its GDPval-AA v2 score increased from 642 to 1,140.
This creates a useful routing architecture. Extracting fields from a known form may need minimal thinking on Flash-Lite; reconciling ambiguous evidence may justify 3.6 Flash. Routing should respond to observed difficulty, not only a developer’s initial guess, and uncertain outputs should escalate visibly.
Flash Cyber has a narrower role: validating and remediating code vulnerabilities. Google is limiting access to governments and vetted partners because the same capability could produce offensive exploit code. Inside its CodeMender agent, several model instances work in parallel and cross-check findings before a human approves the remediation report.
Restricted distribution is a governance choice, not proof of safety. Users still need isolated testing, change review and evidence that a proposed patch resolves the flaw without introducing another.
Completed tasks are the right cost benchmark
A lower list price can be misleading if a model retries, generates excessive reasoning or requires substantial human correction. Teams should record the total tokens, latency, tool failures and review effort for each completed workflow.
Caching repeated context and setting appropriate thinking levels can matter as much as model selection. High-volume agents should have spending ceilings and alerts for sudden increases in output length or looping behaviour.
Quality gates belong at the task boundary. Document agents need traceable citations, coding agents need tests, and computer-use systems need confirmation before irreversible actions. A faster model can amplify mistakes as efficiently as correct work.
Google’s three-model approach acknowledges that enterprise agents are not one workload. Reasoning, throughput and cybersecurity require different combinations of capability and control. The opportunity is not simply to replace one expensive model with a cheaper one. It is to build a measured routing system where each step receives the least costly model that can complete it reliably—and where the application knows when that model has not succeeded.
