For the past year the default logic was simple: bigger model means better model. Google just flipped that. Instead of another giant flagship, it split Gemini into three "speed tiers" so different jobs use different models.
If you run a 24/7 customer-service agent, the monthly token bill is real money. What you want there is not "smartest" but "smart enough, cheap, fast, and not choking the loop." All three lightweight models target agentic workflows.
Three models, three jobs
- Gemini 3.6 Flash: the general workhorse for coding, knowledge work, and multimodal. GA since August 2, priced at $1.50 per million input tokens and $7.50 per million output tokens.
- Gemini 3.5 Flash-Lite: the fastest, cheapest tier, rated at 350 output tokens per second, built for high-throughput repetitive tasks, and wired directly into Google Search.
- Gemini 3.5 Flash Cyber: the security variant for finding and patching vulnerabilities, but limited to a pilot for governments and trusted partners. Not publicly available.
The numbers in plain terms
3.6 Flash emits about 17% fewer output tokens than the previous 3.5 Flash. Sounds small? An agent task often runs through fifteen or twenty loop steps, and shaving over a tenth off each one adds up to a visibly smaller monthly bill. Pair that with Flash-Lite's 350 tokens per second, and an automated "check weather, compare prices, place order" flow responds fast enough that users barely notice the pause.
The model race has turned from "who has the most parameters" to "who gets more done per dollar." For small teams running agents, that is a real win.
Reality check
The most interesting one, Flash Cyber, is locked behind a government pilot, so "AI that patches its own vulnerabilities" stays in a glass case for now. Robotics ER 2 is only public preview, still waiting on infrastructure, pricing, and proof cases. There is also the awkward contrast: Gemini 3.5 Pro, which we covered earlier, slipped its launch over coding shortfalls and is still in partner testing, while the lightweight line shipped first.
Also note: Google deprecated the sampling parameters temperature, top_p, and top_k this round. Old code that hard-codes them will need migration.
What this means for you
Small teams doing overseas business and running agent flows: 3.6 Flash's pricing puts "run automation on a frontier model" within easy math. Instead of blindly calling the most expensive flagship, match the tier to the task (Flash for heavy work, Flash-Lite for pipeline work) and you cut cost directly. Hardware and robotics builders: ER 2's early-access waitlist is open now if you want a seat.
Bottom line: Google is not showing muscle here, it is laying pipe. Only when models are plentiful enough to plug in on demand does the agent business actually work.