Google has released Gemini 3.6 Flash alongside Gemini 3.5 Flash-Lite and a security-tuned Gemini 3.5 Flash Cyber variant, tightening its lineup of production-grade models built for agentic workflows. Announced on July 21, 2026, the trio is designed to lower token cost, cut latency, and hand developers a security-focused model that will ship inside Google DeepMind's CodeMender under a limited-access government-and-partners pilot.
Gemini 3.6 Flash: A Cheaper, Sharper Workhorse
Google's headline number for 3.6 Flash is 17% fewer output tokens versus 3.5 Flash on the Artificial Analysis Index, with pricing dropped to $1.50 per million input tokens and $7.50 per million output tokens. Benchmark gains are broad: DeepSWE coding jumps from 37% to 49%, MLE Bench from 49.7% to 63.9%, OSWorld-Verified computer use from 78.4% to 83.0%, and GDPval-AA v2 knowledge-work scores from 1349 to 1421. Context window is 1 million tokens with a 64K output ceiling, native multimodal input, thinking controls and Computer Use are built in, and the knowledge cutoff advances to March 2026.
Flash-Lite: 350 Tokens/s at $0.30 In / $2.50 Out
Gemini 3.5 Flash-Lite is positioned as the fastest 3.5-class model at 350 output tokens/s per Artificial Analysis, priced at $0.30 per million input tokens and $2.50 per million output tokens. Google says Flash-Lite outperforms the prior Gemini 3.1 Flash-Lite across agentic workloads and, on many agentic and coding evals, edges past larger 3-class Flash — SWE-Bench Pro 54.2% vs. 49.6% and OSWorld-Verified 74.0% vs. 65.1% versus 3 Flash. Terminal-Bench 2.1 climbs from 31% to 54%. Computer Use ships as a built-in tool, and developers can dial thinking levels from minimum to high depending on latency vs. reasoning tradeoffs.
3.5 Flash Cyber in CodeMender
The most restricted release is Gemini 3.5 Flash Cyber — a specialized model fine-tuned for detecting, validating and patching security vulnerabilities. It powers CodeMender, DeepMind's AI agent for code security, and reaches competitive frontier-level performance on the CyberGym benchmark. Google is deliberately gating access: 3.5 Flash Cyber will ship exclusively via CodeMender for governments and trusted partners, framed as a defensive tool for critical-vulnerability triage rather than a general-purpose release.
What's Next: 3.5 Pro and Gemini 4
Alongside today's releases, Google says Gemini 3.5 Pro is in partner testing and will be broadly available "as soon as it's ready." Behind it, the DeepMind team has "started our most ambitious pre-training run yet, for Gemini 4." 3.6 Flash is available now via the Gemini API in Google AI Studio, in Google Antigravity and Android Studio, in Gemini Enterprise for enterprise deployments, and through the Gemini app; 3.5 Flash-Lite is also rolling out in Google Search. Frontier Safety safeguards for CBRN and cyber-offense have been strengthened for 3.6 Flash, with Google emphasizing reduced jailbreak susceptibility while cutting refusals on beneficial uses. The release lands the same week Google is also pushing forward on Gemini 3.5 Pro's 2M-token push and its cyber-defense agent stack that competitors like ReliaQuest and OpenAI are also chasing.
Reporting based on coverage from Google Blog, 9to5Google, GitHub, Droid Life and Apidog.