Google released three new Gemini models on July 21: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and a security focused Gemini 3.5 Flash Cyber. Alongside the launch, the company said Gemini 3.5 Pro remains in testing with selected partners and confirmed that pretraining has begun for Gemini 4, its next major model generation.
The announcement reflects a shift in how frontier AI companies compete. Rather than leading with the single most powerful model, Google is emphasizing cost, speed and reliability for the real world workloads that developers and businesses run at scale.
Gemini 3.6 Flash: fewer tokens, lower cost
Google describes Gemini 3.6 Flash as its main fast model for both everyday and complex tasks. The company says it improves on Gemini 3.5 Flash in programming, data analysis and understanding of documents, charts and other mixed content, while being more economical to run.
According to Google, independent evaluation by Artificial Analysis found that Gemini 3.6 Flash used about 17 percent fewer output tokens than its predecessor to complete the same work. In plain terms, the model finishes many tasks with fewer words and fewer steps. That matters for AI agents, which do not produce a single answer but work through multiple stages, call tools and keep going until a result is reached, consuming tokens at every step.
Pricing is set at $1.50 per million input tokens and $7.50 per million output tokens. Google argues that lower token consumption combined with these rates can make the total cost of an agentic task lower than it was with the previous Flash model.
On benchmarks, Google reports that Gemini 3.6 Flash scored 49 percent on the DeepSWE coding test, up from 37 percent for Gemini 3.5 Flash. On OSWorld-Verified, which measures a model’s ability to operate a computer to complete tasks, it reached 83 percent compared with 78.4 percent for the earlier model. These are company reported figures.
Gemini 3.5 Flash-Lite for high volume work
Gemini 3.5 Flash-Lite is built for very fast, low cost use. Google positions it for business workloads that involve large numbers of requests each day, such as extracting information from documents, summarizing content, working across languages and handling small subtasks inside agent pipelines.
Google calls it the fastest model in the 3.5 series, citing Artificial Analysis measurements of roughly 350 output tokens per second. It is priced at $0.30 per million input tokens and $2.50 per million output tokens.
The company also says Flash-Lite is substantially more capable than the previous Lite generation. In some programming and practical task tests, Google reports that it outperformed not only Gemini 3.1 Flash-Lite but in places the older Gemini 3 Flash as well, a sign that lower cost models are now handling work that recently required larger ones.
A restricted model for security teams
The third model, Gemini 3.5 Flash Cyber, is not aimed at general users. Google says it is tuned to find, test and help fix security vulnerabilities in software, and the company is using it with CodeMender, its security system in which multiple AI agents work together on a security issue.
Because of the sensitivity of that capability, Google is not releasing Flash Cyber publicly. In the first phase it will be offered through CodeMender to governments and trusted partners under a limited trial program. The stated goal is to help defensive teams find and patch dangerous flaws before attackers do.
Availability and what comes next
Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are available now through the Gemini API, Google AI Studio and the Gemini app, among other platforms. Google said Flash-Lite is also being integrated into Google Search.
Two forward looking items drew attention. Google confirmed that Gemini 3.5 Pro is in testing with selected partners and will be released more widely once ready, without giving a date. And for the first time, the company said work on Gemini 4 is underway, describing the start of its largest pretraining run to date. No details on capabilities or timing were given.
Taken together, the releases show Google balancing quality against cost with Gemini 3.6 Flash, targeting speed and volume with Flash-Lite, and moving toward domain specific models with Flash Cyber, while continuing to build the next flagship generation in parallel.
