Google Ships Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Google released three Gemini models on July 21, 2026: Gemini 3.6 Flash for coding and knowledge work, Gemini 3.5 Flash-Lite for high-throughput, low-latency agentic tasks, and Gemini 3.5 Flash Cyber, a security-specialized model available only to governments and trusted partners (Google blog). The interesting move is at the cheap end: Google is pricing agentic throughput aggressively while its flagship 3.5 Pro is still in partner testing.

Key Takeaways

• Gemini 3.6 Flash scores 49% on DeepSWE (vs 37% for 3.5 Flash), 63.9% on MLE Bench (vs 49.7%), and 83.0% on OSWorld-Verified (vs 78.4%), at $1.50/M input and $7.50/M output tokens (Google blog).
• 3.6 Flash uses 17% fewer output tokens than 3.5 Flash on average, and up to 65% fewer on DeepSWE, so the sticker price understates the effective savings on agentic work.
• Gemini 3.5 Flash-Lite streams at 350 output tokens per second and costs $0.30/M input, $2.50/M output. It hits 54% on Terminal-Bench 2.1 (vs 31% for 3.1) and beats 3 Flash on SWE-Bench Pro (54.2% vs 49.6%).
• Gemini 3.5 Flash Cyber runs inside Google's CodeMender agent to find and fix vulnerabilities, with competitive frontier performance on CyberGym. It is a limited-access pilot for governments and trusted partners only.
• Gemini 3.5 Pro remains in partner testing, and Google says Gemini 4 pre-training is underway, described as its most ambitious run yet.

The Three Models at a Glance

Why the Token-Efficiency Number Matters Most

The 17% output-token reduction is the number to anchor on, because agentic coding bills scale with output tokens, not with list price alone. On DeepSWE specifically, Google reports up to a 65% output reduction, meaning 3.6 Flash can be dramatically cheaper per completed task than its predecessor even before you compare per-token rates. If you run high-volume agent loops, benchmark cost per finished task, not cost per million tokens.

Flash-Lite's positioning is just as pointed: 350 tokens per second at $0.30/$2.50 is a direct play for the tool-calling and orchestration tier where fast, cheap models like GPT-5.6 Luna compete. Its 54% on Terminal-Bench 2.1 is a striking claim for a Lite-tier model on an agentic terminal benchmark.

Flash Cyber Is a Different Kind of Release

Gemini 3.5 Flash Cyber is not a product you can buy; it is a controlled deployment. Google is piloting it exclusively with governments and trusted partners inside CodeMender, where multiple agents cooperate to produce combined vulnerability reports. This mirrors the pattern OpenAI set with GPT-Red, its internal-only red-teaming model: frontier labs are building offensive-capable security models and deliberately keeping them off the public shelf.

How to Use This Update

If you build on the Gemini API, run one real workload through 3.6 Flash this week and measure output tokens against your current model; the efficiency gain is the kind of thing that only shows in your own traces. If you run high-volume agent pipelines, put Flash-Lite in your router for tool-calling and formatting steps and keep the heavier model for planning. And if you are on Google's stack via the Interactions API, check model-name pinning before you swap defaults.