Google has released two new entries in its Gemini lineup, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, while offering a brief look at what comes next. The updates arrive as the company continues to iterate on models that balance speed, cost, and capability in a crowded field where incremental gains often matter more than dramatic leaps.
Gemini 3.6 Flash builds on the version shown at I/O earlier this year. Google says it draws on feedback collected since May and focuses on using fewer tokens across tasks. Independent measurements show it produces roughly 17 percent fewer output tokens than Gemini 3.5 Flash. It also takes fewer reasoning steps and tool calls when handling multi-step work. Pricing has been adjusted downward to $1.50 per million input tokens and $7.50 per million output tokens, a noticeable drop from the previous output rate of $9. In coding tests the model posts higher scores on DeepSWE, rising from 37 percent to 49 percent, and on MLE Bench, moving from 49.7 percent to 63.9 percent. Knowledge-work benchmarks reach 1421 on GDPval-AA compared with 1349 before, while computer-use results on OSWorld-Verified climb from 78.4 percent to 83 percent. The knowledge cutoff has finally moved forward from January 2025 to March 2026, closing a gap that had begun to feel dated.
Alongside it sits Gemini 3.5 Flash-Lite, aimed at high-volume, low-latency jobs such as agentic search and document handling. Priced at $0.30 per million input tokens and $2.50 per million output tokens, it improves on the earlier 3.1 Flash-Lite released in March. Gains appear in Terminal-Bench 2.1, where scores rise from 31 percent to 54 percent, in long-context evaluation on GDM-MRCR v2 from 60.1 percent to 72.2 percent, and in real-world task scores on GDPval-AA v2 from 642 to 1140. It also edges past the prior 3 Flash model on SWE-Bench Pro and OSWorld-Verified.
A third variant, Gemini 3.5 Flash Cyber, targets security work. Built on the efficiency of the Flash series, it is intended to detect, validate, and patch code vulnerabilities at scale. Access remains limited to governments and selected partners through a pilot program that pairs the model with Google’s CodeMender tool. The restricted rollout reflects ongoing caution about dual-use risks in automated vulnerability discovery.
Both 3.6 Flash and 3.5 Flash-Lite are live in the Gemini app today, with the lighter model also headed to Search. Developers can reach them through Google’s Antigravity interface, AI Studio, and Android Studio. Gemini 3.5 Pro remains in partner testing with no firm public date. In parallel, the DeepMind team has already begun what it calls its most ambitious pre-training run yet for Gemini 4.
These releases fit a familiar pattern in the current AI cycle: steady efficiency improvements, selective benchmark advances, and careful pricing adjustments rather than sudden capability jumps. Whether the gains translate into everyday reliability for developers and enterprises will depend on real-world use beyond the controlled test suites.
