Google launches Gemini 3.6 Flash, a faster 3.5 Flash-Lite, and a cyber-only model

智东西 · wechat · 2026-07-22

Google released three new Gemini models focused on Agent efficiency: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber.

Gemini 3.6 Flash

Built for coding, knowledge work, and multimodal tasks.
Uses 17% fewer tokens than 3.5 Flash on ArtificialAnalysis’ Intelligence Index, and up to 65% fewer output tokens on DeepSWE.
Costs $1.50 per million input tokens and $7.50 per million output tokens.
Scores 49% on DeepSWE v1.1, 63.9% on MLE-Bench, 1421 on GDPval-AAv2, and 83.0% on OSWorld-Verified.
Google says it is especially strong on speed and lower Agent run cost.

Gemini 3.5 Flash-Lite

Google’s fastest and cheapest Flash-Lite model, aimed at low-latency, high-throughput Agent search and document work.
Reaches 350 output tokens per second.
Priced at $0.30 per million input tokens and $2.50 per million output tokens.
Improves over 3.1 Flash-Lite across Terminal-Bench 2.1, long-context retrieval, and GDPval-AAv2.

Gemini 3.5 Flash Cyber

A security-focused model for finding, validating, and fixing vulnerabilities.
Not open to general developers yet; it will be trialed with CodeMender and trusted partners.
Google says its performance is competitive with frontier cyber models, but it is being withheld because of dual-use risk.

Google also says Gemini 3.5 Pro is still in partner testing, and Gemini 4 has entered the company’s largest pretraining run to date.

More from coding & agent