Gemini 3.6 Flash Beats Its Predecessor on Coding While 3.5 Flash-Lite Chases Speed

What happened

Google shipped three new Gemini models on July 21: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. Google frames the family around efficiency, latency, and reliability for building AI agents at scale, and each model targets a different point on that tradeoff.

Gemini 3.6 Flash is the capability upgrade. Google says it delivers better coding, knowledge work, and multimodal performance than 3.5 Flash, while using 17% fewer output tokens for the same work. On DeepSWE, a coding benchmark, Google reports up to a 65% improvement. On OSWorld-Verified, a test of completing tasks inside a real desktop environment, it scores 83.0% against 78.4% for 3.5 Flash. It costs $1.50 per million input tokens and $7.50 per million output tokens, and is available now through the Gemini API in Google AI Studio and Android Studio, inside Google Antigravity, in the Gemini Enterprise Agent Platform and app, and in the consumer Gemini app.

Gemini 3.5 Flash-Lite goes the other way. Google calls it its fastest, most cost-effective 3.5-class model, and Artificial Analysis measured it running at 350 output tokens per second. Google says it beats Gemini 3 Flash on SWE-Bench Pro, a coding benchmark, 54.2% to 49.6%, and significantly outperforms the earlier 3.1 Flash-Lite on agentic workflows, tasks where a model calls tools and takes multiple steps toward a goal. It’s priced at $0.30 per million input tokens and $2.50 per million output tokens, roughly a fifth of what 3.6 Flash costs, and reaches the same channels as 3.6 Flash plus a rollout in Google Search.

The third model, Gemini 3.5 Flash Cyber, is fine-tuned to find and fix security vulnerabilities and performs competitively on CyberGym, a cybersecurity benchmark. It isn’t a model most builders can reach. Google is limiting it to governments and trusted partners through a CodeMender pilot program.

Gemini 3.6 Flash also showed up the same day in GitHub Copilot, rolling out to Copilot Pro, Pro+, Max, Business, and Enterprise users across VS Code, Visual Studio, the Copilot CLI, the Copilot cloud agent, the Copilot app, JetBrains, Xcode, and Eclipse. That’s the same model Google just released, reaching builders through a second, familiar interface, billed at provider list pricing under Copilot’s usage-based billing. Business and Enterprise admins need to enable the “Gemini 3.6 Flash Preview” policy in Copilot settings before anyone in their org can select it.

The real decision: which model for which job

Two of these three models are usable by any builder today, and they pull in opposite directions. Gemini 3.6 Flash is the one to reach for when the task itself is hard: a real refactor, a multi-file bug fix, a task with a lot of ambiguity where a wrong first step is expensive. Gemini 3.5 Flash-Lite is the one to reach for when the task is simple but you’re running it a lot: classifying incoming support tickets, extracting fields from documents, powering a chat feature where response time matters more than depth.

Picking wrong shows up as either wasted money or a frustrating agent. Routing volume, low-stakes calls through the pricier, more capable model burns budget for no real gain. Routing genuinely hard work through the cheap, fast model produces more retries and corrections, which can erase the savings and cost you time on top.

The benchmark numbers above are Google’s own reporting. That doesn’t make them wrong, but it does mean they describe Google’s test conditions, not necessarily your workload.

What builders should do next

Don’t take the benchmark deltas as your answer. Take one task you already run regularly with your current model, one with a clear pass or fail, like a bug fix that either passes your existing tests or doesn’t. Run it once against Gemini 3.6 Flash and once against Gemini 3.5 Flash-Lite through the Gemini API in Google AI Studio, keeping the prompt and starting point identical both times.

Compare four things afterward: whether the result actually passed your tests, how many output tokens each run consumed, how long each run took, and how many follow-up corrections you needed before the result was usable. That gives you a real answer for your own workload instead of a borrowed one from Google’s benchmark suite.

If you reach Gemini through GitHub Copilot instead, check your IDE’s model picker first. If Gemini 3.6 Flash isn’t listed yet and you’re on a Business or Enterprise plan, the rollout may still be pending, or your organization’s admin may need to turn on the preview policy before it appears for you.

End of article