TLDRocket
Sign in

Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Google DeepMind Covered by 19 sources

Google shipped three new Gemini models: 3.6 Flash, 3.5 Flash-Lite, and a cybersecurity-focused 3.5 Flash Cyber for finding bugs. They're cheaper and faster than before, and Google's already deep into training Gemini 4.

Based on reporting by Google DeepMind — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Google DeepMind just refreshed its Flash lineup again, barely pausing since the 3.5 Flash launch, and the theme this time is doing more with fewer tokens. Gemini 3.6 Flash is the headline act: Google says it cuts output token usage by 17% compared to 3.5 Flash on the Artificial Analysis Index, and on DeepSWE, a coding benchmark from Datacurve, that efficiency gain jumps to as much as 65%. It's priced at $1.50 per million input tokens and $7.50 per million output tokens, undercutting its predecessor while scoring higher on knowledge-work tests like GDPval-AA v2 and computer-use tasks like OSWorld-Verified, where it hit 83.0% versus 78.4% for 3.5 Flash.

Then there's 3.5 Flash-Lite, built for the grunt work of agentic systems: high-volume search, document parsing, the stuff that needs speed more than depth. Google clocks it at 350 output tokens per second, and at $0.30 per million input tokens and $2.50 per million output tokens, it's aimed squarely at production traffic where cost per call actually matters. What's notable is that Flash-Lite reportedly beats the older, pricier 3 Flash on several benchmarks, including SWE-Bench Pro and OSWorld-Verified, which suggests Google is comfortable letting its cheapest tier eat into its mid-tier's lunch.

The oddest release in the bunch is 3.5 Flash Cyber, a specialized model tucked inside Google's CodeMender agent. It's designed to hunt for and patch software vulnerabilities faster than attackers can exploit them, running multiple agent instances that compile a single vulnerability report. Google isn't opening this one up broadly. It's going out through a limited pilot to governments and select partners only, an acknowledgment that a model good at finding security holes is also a model good at exploiting them.

Buried in the announcement is the bigger tell: Gemini 3.5 Pro is still in partner testing, and Google says it has already started its most ambitious pre-training run yet, for Gemini 4. That's a lot of forward momentum stacked on top of a Flash refresh that's barely a season old. Google is clearly treating Flash as the workhorse tier to keep enterprise agents cheap and fast while the real fight for frontier bragging rights happens elsewhere, in Pro and whatever Gemini 4 turns out to be.

My take — AI-written commentary, not fact-checked reporting

The token-efficiency numbers matter more than the benchmark scores here, because agentic workloads live and die by cost per action, not raw intelligence. What I'd flag is the Cyber model's gated release: it's the right call, but it's also a quiet admission that AI vulnerability discovery has outpaced patching, and closed, permissioned deployment is becoming the default answer to dual-use risk whether open-model advocates like it or not.

Read more about this at: Google DeepMind

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.