.png)
Google DeepMind released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber in CodeMender. The update is aimed squarely at production agent workloads: lower latency, fewer output tokens, and better cost/performance in the Flash tier.
Gemini 3.6 Flash is the new workhorse model in the series. Google DeepMind says it improves coding, knowledge work, and multimodal performance while also reducing output token usage versus Gemini 3.5 Flash. On the Artificial Analysis Index, the reported reduction is 17%; in some coding-heavy benchmarks such as DeepSWE, the reduction is much larger, up to 65%.
That matters because output tokens are often the real bill in agentic systems. If a model reaches the same task completion rate with fewer generated tokens, the gains show up in latency, throughput, and unit economics at the same time. For teams running tool-using agents or long-horizon workflows, the right comparison is not just benchmark score or raw token/s — it is “successful task completions per dollar” and “successful task completions per second.”
The other part of the story is that 3.6 Flash is not being positioned as a narrow specialist. DeepMind is calling it the general-purpose Flash option for coding, knowledge work, and multimodal use cases, which makes it the obvious candidate for replacing 3.5 Flash in production systems that already use the Flash tier as the default model.
Gemini 3.5 Flash-Lite is the fastest and most cost-effective 3.5-class model in the release. DeepMind cites 350 output tokens per second on the Artificial Analysis Index and says it significantly improves agentic workflows compared with earlier Flash-Lite generations.
That makes Flash-Lite the model to evaluate for high-throughput orchestration steps: classification, routing, extraction, simple tool calls, short-form generation, and “good enough” sub-agents that should never dominate the cost profile of the system. The practical question for teams is whether the slightly lower quality ceiling is offset by enough throughput and latency improvement to justify deploying it in places where 3.6 Flash would be overkill.
If you already run cascades or model routers, this is the model to slot into the lowest tier and measure on real traffic. The important metric is not whether it wins a benchmark category; it is whether it preserves workflow success rates when used in the high-volume, low-complexity parts of your pipeline.
The most interesting part of the announcement is not just a model release. It is the framing around Gemini 3.5 Flash Cyber in CodeMender: cybersecurity performance depends on the model plus the agent infrastructure around it.
That is the right way to think about security-oriented AI systems. A cyber model by itself does not solve code security; the agent loop, verification steps, patch application, and feedback mechanisms matter just as much. DeepMind is explicitly treating the model as one component in a larger orchestration stack, which is consistent with how security workflows are actually deployed when they need to be reliable.
For engineers evaluating similar systems, this is a reminder to benchmark the full pipeline, not the LLM in isolation. Measure patch validity, regression rate, review burden, exploitability reduction, and how often the agent produces a safe change versus a plausible but incorrect one. That is especially important in security, where “looks good” is not a useful acceptance criterion.
The report also says the Cyber combination delivers competitive frontier performance. The exact technical details of the model and agent setup are still thin in the release material, so the main actionable signal is architectural: specialized models work best when they are paired with equally specialized control logic.
DeepMind also says Gemini 3.5 Pro is testing with partners and will be broadly available once it is ready. That places the release in a clear product hierarchy: Flash for efficiency and scale, Pro for higher-end capability when it becomes generally available, and specialized variants like Flash Cyber for constrained domains.
DeepMind additionally says work has already started on Gemini 4, with its most ambitious pre-training run so far. That is interesting as roadmap context, but it does not change near-term deployment decisions. For now, the practical choice is between 3.6 Flash and 3.5 Flash-Lite depending on whether your bottleneck is quality or throughput, and whether your system needs a cyber-specialized path.
For production teams, the actionable move is straightforward: evaluate 3.6 Flash as the default replacement for 3.5 Flash, use 3.5 Flash-Lite where latency and cost dominate, and treat the CodeMender announcement as evidence that the best security agents will be system designs, not single-model demos.
