Google DeepMind has released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and a restricted cybersecurity model called Gemini 3.5 Flash Cyber, expanding its selection of lower-cost, production-oriented AI systems while leaving the anticipated Gemini 3.5 Pro out of the launch.
The company is positioning Gemini 3.6 Flash as its general-purpose workhorse for coding, knowledge work and multimodal applications. Google says the model can use up to 17% fewer tokens than Gemini 3.5 Flash, reducing the cost of completing some tasks. It is priced at $1.50 per million input tokens and $7.50 per million output tokens and supports a context window of one million tokens.
Google reported improvements over Gemini 3.5 Flash across several evaluations. Its score on the DeepSWE software engineering benchmark rose from 37% to 49%, while its MLE-Bench result increased from 49.7% to 63.9%. On OSWorld-Verified, an evaluation of agents interacting with computer environments, the score moved from 78.4% to 83%. Benchmark results can vary with testing configurations and do not necessarily reflect performance in every production workload.
The release also makes Computer Use available as a built-in client-side tool in the Gemini API and Gemini Enterprise. The capability is intended to let applications direct models to interact with software interfaces, including browsers and desktop environments. Google has also added safeguards aimed at reducing the risk of the models being misused for cyberattacks or work involving chemical, biological, radiological and nuclear threats.
Gemini 3.5 Flash-Lite targets applications that need high throughput and low latency at a lower price. It costs $0.30 per million input tokens and $2.50 per million output tokens. Google says it surpasses older Flash models on several coding and agent evaluations. The model is designed for large-scale services where response speed and operating cost may matter more than the maximum reasoning performance available from a frontier model.
The third release, Gemini 3.5 Flash Cyber, is based on Gemini 3.5 Flash but has been tuned to identify and help repair software vulnerabilities. Google has incorporated it into CodeMender, a DeepMind security agent that can deploy multiple specialized subagents in parallel before combining their findings into a report.
In Google’s testing, Flash Cyber scored 83.2% on the CyberGym benchmark. The company also said the model identified 55 confirmed, distinct findings while examining commits to the V8 JavaScript engine, compared with 47 for the standard Gemini 3.5 Flash. Ten of the Flash Cyber findings were not detected by the other models included in that test.
Because vulnerability research tools can assist both defenders and attackers, Flash Cyber will not receive a general public release. Google plans to provide it only to governments and trusted partners through a limited pilot using CodeMender. A broader preview of the CodeMender agent is available through the Gemini Enterprise Agent Platform, but that version uses Google’s standard Gemini models. It supports C/C++, Go, Java, Python, Ruby, Rust and TypeScript.
The absence of Gemini 3.5 Pro is notable because Pro models generally occupy the high-capability tier of Google’s lineup, with an emphasis on demanding reasoning and coding tasks. Flash systems instead prioritize speed, efficiency and lower inference costs, characteristics that are especially relevant to companies running AI agents repeatedly or at high volume.
Google previously indicated that Gemini 3.5 Pro was being used internally and was expected to arrive after the May introduction of Gemini 3.5 Flash. Bloomberg reported last week, citing internal delays, that the company had struggled to get the model to meet its performance targets. Google DeepMind product lead Logan Kilpatrick said the Pro model is now being tested with partners and that the company hopes it will arrive soon, without giving a release date.
Meanwhile, Google has begun pretraining Gemini 4 in what the company describes as its most ambitious training run to date. Until Gemini 3.5 Pro is ready, Google’s latest public update remains centered on the Flash family and the practical requirements of deploying agents: lower latency, predictable operation and reduced inference costs.
Sources: TechCrunch AI, The Decoder