Google expanded Gemini's Flash lineup with 3.6, Flash-Lite and Cyber models

Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite in a single July 21 announcement. The same post introduced Gemini 3.5 Flash Cyber as a specialized defensive-security model that would become available soon, exclusively to governments and trusted partners through CodeMender in a limited-access pilot. Google positioned 3.6 Flash as the general workhorse and Flash-Lite for high-volume and latency-sensitive tasks. The announcement emphasizes different efficiency and task profiles rather than one model replacing the entire line.

Model routing became a product decision

A broader family lets an application choose a smaller or specialized model, but routing adds a new source of regressions. The cheapest qualifying model for classification may not be the right one for a multi-step tool workflow. A security-tuned model may generate findings that still require validation and safe patch review.

Avoid a router that learns only from synthetic benchmark labels. Start with explicit task classes and route rules that operators can inspect. Log the selected model and fallback, then compare correction rate, latency and total workflow cost on production-shaped examples.

Pin, evaluate and stage each variant

Build a separate acceptance suite for each path. Flash-Lite tests should include high concurrency and ambiguous short inputs. General Flash tests should include long context and tool failures. Cyber tests should include false positives, incomplete repositories and patches that appear plausible but break behavior.

Treat model identifiers and routing policy as versioned configuration. Canary changes, retain the prior route and watch downstream outcomes rather than only API error rates. Google’s release gives builders more explicit choices; reliable adoption depends on proving where each choice stops being good enough.