Google splits its Flash line between general agents, high-volume work, and restricted cyber use

Google announced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber on July 21. The first two are generally available in the Gemini API and positioned for production use; Cyber is a separate security-focused model paired with CodeMender and is not a broadly available API option. Google says the Cyber pilot will initially be limited to governments and trusted partners, an important boundary that separates a product announcement from evidence of general access.

The practical split is unusually explicit in the API documentation. Gemini 3.6 Flash is the higher-capability general model for coding, multimodal reasoning, and multi-step agent workflows, while 3.5 Flash-Lite is the lower-cost, high-throughput option for extraction, structured parsing, and subagent execution. Both expose a one-million-token context window, up to 64,000 output tokens, thinking, and built-in tools including Computer Use, so the choice is not simply between a feature-complete model and a stripped-down batch endpoint.

Against Gemini 3.5 Flash, Google lists the same $1.50 per million input tokens for 3.6 Flash but a lower $7.50 output-token price versus $9.00. Google also reports lower output-token use and better results on selected coding, knowledge-work, and computer-use evaluations. Those are supplier claims and benchmark-specific results, not an independently normalized comparison with competing models, but the combination of token use and output pricing is a concrete operational metric for teams whose agents spend heavily on long responses and repeated tool loops.

Migration requires more than changing a model ID. Google documents that the new models deprecate temperature, top_p, and top_k, reject prefilled model turns, and change how developers should configure reasoning. Existing applications that rely on those request shapes can see behavioral changes or validation errors, so a staged integration test is more useful than a blind replacement. Flash-Lite also defaults to minimal thinking; Google recommends raising the thinking level for autonomous planning and multi-step tool work, which means its advertised throughput should be evaluated against the reasoning setting an application actually needs.

For technical buyers, the announcement is best read as a routing decision rather than a single-model upgrade. Use 3.6 Flash where a general agent needs stronger multimodal or coding behavior and where lower output cost can compound across many turns; use 3.5 Flash-Lite where volume, extraction, and inexpensive subagents dominate. Treat Flash Cyber as a restricted security pilot until its access terms change, and validate Google’s performance claims on the workload, tool configuration, and safety constraints that apply to the deployment.