Google released three new AI models on July 21, 2026: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The releases emphasize operational efficiency and specialization. Flash 3.6 cuts output token usage by 17 percent compared to its predecessor. Flash-Lite runs at 350 tokens per second with pricing that needs comparison against the workload being considered. Flash Cyber is a security specialist paired with a code scanning agent called CodeMender.
These releases offer different options for teams balancing speed, cost, and the demands of a particular job. For example, a customer service workload might target a 200 millisecond response and 10,000 concurrent sessions at a fraction of a cent per interaction. These hypothetical requirements would need their own performance test.
According to Artificial Analysis, an independent benchmarking firm, Flash-Lite delivers 350 output tokens per second. A team still needs to test end-to-end latency and cost on its own tasks; token throughput alone does not establish either.
On certain benchmarks such as DeepSWE, Flash 3.6 reportedly uses up to 65 percent fewer tokens to produce the same result. That task-specific claim has a different scope from the release’s 17 percent output-token reduction. Pricing, retry rates, and tool calls also affect the cost of a completed task.
A model comparison is useful when the team knows the workload it needs to support. Test the candidates on that work and include failures that require a retry or human review.
The workflow defines the model’s job, its data access, and the permissions for any action it can take. Monitoring and feedback need an owner, including someone who can handle a wrong result.
The reported 17 percent reduction concerns output-token usage. To estimate savings, apply the relevant prices to a representative task and include the surrounding review work. Keep the workflow steady during comparison so a process change does not obscure the model’s contribution.
The same throughput can matter differently in a sidebar chatbot and a hypothetical company running 40 agents across procurement, compliance, and customer operations. At thousands of tasks per hour, concurrency limits and review capacity need testing alongside output speed.
Flash Cyber and CodeMender pair a specialist model with a code-scanning agent. A security team should evaluate that pairing on its own review tasks, with the same access and approval constraints it would require of another tool.
For a team considering these models, retain a record of task quality, latency, and total cost under the same test conditions. That gives the next release a concrete comparison point. A lower token bill is useful when the model still completes the job within the workflow’s quality and safety requirements.