Google separates its Flash line by agent workload, cost, and access

Google’s July 21 announcement presents Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber as three different deployment choices rather than a single version refresh. The accompanying Gemini API model directory labels 3.6 Flash and 3.5 Flash-Lite as stable, while Flash Cyber does not appear in that public stable-model list. The documented distinction is consequential: Google describes the first two as broadly usable developer and enterprise options, but describes Cyber through a controlled CodeMender path.

For Gemini 3.6 Flash, Google’s central claim is that less output and fewer agent steps can reduce the cost of completing multi-step work. It lists pricing of $1.50 per million input tokens and $7.50 per million output tokens, and cites a 17% reduction in output-token use versus Gemini 3.5 Flash on the Artificial Analysis Index. Google also reports gains over 3.5 Flash on DeepSWE, MLE Bench, OSWorld-Verified, and GDPval-AA v2. Those figures are useful deployment hypotheses, but they are provider-reported comparisons in the materials reviewed here, not independently replicated results.

Gemini 3.5 Flash-Lite targets a different economic profile. Google lists $0.30 per million input tokens and $2.50 per million output tokens, and cites 350 output tokens per second from Artificial Analysis. Its announcement compares the model with 3.1 Flash-Lite on Terminal-Bench 2.1, long-context retrieval, and GDPval-AA v2, while also describing configurable thinking levels and a built-in computer-use tool. For teams running extraction, search, routing, or subagent work at volume, the relevant question is likely throughput at an acceptable error rate, not whether Flash-Lite can replace a higher-capability coordinator on every task.

Flash Cyber is the important limitation in the family announcement. Google says it is based on 3.5 Flash and tuned for finding and fixing software vulnerabilities, but says access will be exclusive to governments and trusted partners through a limited CodeMender pilot. That means its benchmark and agent-orchestration claims should not be read as evidence of a generally available API choice. For the two public stable model names, Google’s API documentation also advises production users to target a specific stable identifier; it cautions that preview identifiers can have tighter limits and are subject to deprecation. The verified record is therefore the provider’s documented positioning and access terms, while independent testing of cost, quality, and safety tradeoffs remains outstanding.