Google has introduced Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, two models designed to improve the speed, cost and reliability of production AI agents. The company also unveiled Gemini 3.5 Flash Cyber, a specialized cybersecurity model that will operate through Google’s CodeMender security agent.
Gemini 3.6 Flash is positioned as Google’s primary workhorse model for coding, knowledge work and multimodal applications. According to Google, the model requires 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index while using fewer reasoning steps and tool calls for multi-step workflows.
Reducing token usage can lower the cost of running AI agents, particularly when businesses are processing large volumes of requests. Google priced Gemini 3.6 Flash at $1.50 per one million input tokens and $7.50 per one million output tokens, below the output pricing of its predecessor.
Google reported that Gemini 3.6 Flash scored 49% on the DeepSWE coding benchmark, compared with 37% for Gemini 3.5 Flash. It also achieved 63.9% on MLE Bench for machine learning research, compared with 49.7% for the earlier model.
The model’s computer-use performance reached 83% on OSWorld-Verified, up from 78.4% for Gemini 3.5 Flash. Computer use allows an AI agent to interact with software interfaces and complete tasks across applications rather than relying exclusively on text responses.
Gemini 3.6 Flash also improved on Google’s cited knowledge-work evaluations. It scored 1,421 on GDPval-AA v2, compared with 1,349 for Gemini 3.5 Flash, and customers have tested the model for document parsing, chart analysis, data interpretation and report drafting.
Google is shipping Gemini 3.6 Flash with expanded safeguards related to chemical, biological, radiological and nuclear risks and offensive cybersecurity misuse. The company said the protections are designed to increase resistance to jailbreak attempts while minimizing unnecessary refusals for legitimate applications.
Gemini 3.5 Flash-Lite is intended for high-volume workloads where latency and cost are particularly important. Google said the model generates approximately 350 output tokens per second based on Artificial Analysis testing.
The model is priced at $0.30 per one million input tokens and $2.50 per one million output tokens. Developers can adjust its thinking level to prioritize low-cost, rapid execution or enable additional reasoning for more complex agent and subagent workflows.
Google reported that Gemini 3.5 Flash-Lite scored 54% on Terminal-Bench 2.1, compared with 31% for Gemini 3.1 Flash-Lite. It also reached 72.2% on GDM-MRCR v2 for long-context performance and 1,140 on GDPval-AA v2, compared with 60.1% and 642, respectively, for the earlier model.
On some coding and agentic evaluations, Gemini 3.5 Flash-Lite also outperformed Gemini 3 Flash. Google cited scores of 54.2% versus 49.6% on SWE-Bench Pro and 74% versus 65.1% on OSWorld-Verified.
Gemini 3.5 Flash Cyber is a version of Gemini 3.5 Flash fine-tuned to detect, validate and repair software vulnerabilities. Within CodeMender, multiple Flash Cyber agents can collaborate to analyze security issues and produce a consolidated report.
Because cybersecurity models can be used for both defensive and malicious purposes, Google plans to limit Gemini 3.5 Flash Cyber to governments and trusted partners through an upcoming CodeMender pilot. The company said the restricted deployment is intended to help defenders identify vulnerabilities while reducing the risk of broader misuse.
Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are available through the Gemini API in Google AI Studio and Android Studio. They are also available through Google’s Gemini Enterprise Agent Platform and the Gemini app, while Flash-Lite is beginning to roll out in Google Search.
Gemini 3.6 Flash is additionally available through Google Antigravity and the Gemini Enterprise application. Google is separately testing Gemini 3.5 Pro with partners and has begun its most ambitious pre-training run to date for Gemini 4.
