Google launched three new Gemini models on July 21, and the pitch is not about intelligence — it is about cost. Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber all target the same bottleneck: AI agents are too expensive to run at production scale, and Google thinks it can fix that with better token efficiency.
The releases come a week after Gemini 3.5 Pro leaked ahead of schedule and Google delayed its full rollout. Today’s models fill the gap with a workhorse (3.6 Flash), a speed demon (3.5 Flash-Lite), and a security specialist (3.5 Flash Cyber) — each aimed at a different slice of the agentic market.
🔍 THE BOTTOM LINE
Google’s model strategy has shifted from “smartest model wins” to “cheapest agent wins.” A 17% token reduction sounds modest until you multiply it across millions of API calls per day — that is the difference between a demo and a product. The cybersecurity model, restricted to governments, is the more interesting signal: Google is carving out vertical-specific models before competitors realise verticals matter.
What 3.6 Flash Actually Delivers
The headline number: 3.6 Flash uses 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, and in some coding benchmarks like DeepSWE by Datacurve, the reduction reaches 65%. At $1.50 per million input tokens and $7.50 per million output tokens, that is a direct cost cut for anyone running agentic workflows.
The performance gains are real but incremental: DeepSWE scores up from 37% to 49%, MLE Bench from 49.7% to 63.9%, and OSWorld-Verified computer-use from 78.4% to 83.0%. These are not generational leaps. They are the kind of improvements that matter when you are running 10,000 agent loops per hour and a 5% accuracy bump means 500 fewer retries.
Google says 3.6 Flash ships with enhanced Frontier Safety safeguards against chemical, biological, radiological, nuclear, and cyber-offence misuse — while claiming to minimise refusals for legitimate use. That last claim is the one to watch. Every model maker says it; few deliver.
3.5 Flash-Lite: The Throughput Play
At 350 output tokens per second and $0.30 per million input tokens, 3.5 Flash-Lite is Google’s answer to the question: what if your agent does not need to be smart, just fast? The model outperforms 3.1 Flash-Lite on Terminal-Bench 2.1 (54% vs 31%), long-context retrieval (72.2% vs 60.1%), and real-world task execution (1140 vs 642 on GDPval-AA v2).
More striking: on several agentic and coding evals, 3.5 Flash-Lite outperforms the older 3 Flash model — including SWE-Bench Pro (54.2% vs 49.6%) and OSWorld-Verified (74.0% vs 65.1%). A cheaper, faster model beating a supposedly smarter one is the pattern that defined the original Flash line, and it is the pattern that keeps pressure on OpenAI and Anthropic to justify their premium pricing.
3.5 Flash Cyber: The Restricted Model
The most unusual release is 3.5 Flash Cyber, a cybersecurity-specialised model available only to governments and trusted partners through Google’s CodeMender agent. It is fine-tuned for finding and patching code vulnerabilities, and Google is deliberately limiting access.
This is a departure from the open-everyone API model that defines the frontier AI market. Google is saying: some models are too dangerous to sell to anyone with a credit card, but useful enough that governments should get them first. The competitive implication is that cybersecurity becomes a vertical where Google has a product nobody else can match — because nobody else has built a restricted-distribution model for it. As we noted in our coverage of AI killing cybersecurity competitions, the offensive-defensive AI security race is already reshaping the field.
NZ Angle
New Zealand developers building agentic workflows on Google Cloud get the cost reduction immediately — 3.6 Flash and 3.5 Flash-Lite are available via the Gemini API and Google AI Studio today. For NZ government agencies interested in 3.5 Flash Cyber, the restricted-access model raises the same question every allied-nation security product does: does NZ qualify as a “trusted partner,” and if so, through what procurement path? The answer is likely yes via Five Eyes, but the process will not be fast.
❓ FAQ
Are these models available in New Zealand? Yes. 3.6 Flash and 3.5 Flash-Lite are available globally through the Gemini API and Google AI Studio. 3.5 Flash Cyber is restricted to governments and trusted partners.
How does 3.6 Flash compare to Claude or GPT for coding? Google’s benchmarks show improvements over its own 3.5 Flash, but direct comparisons with Claude or GPT models are not in the release. The Artificial Analysis Index provides independent rankings.
What is Gemini 3.5 Pro, and when is it coming? Google said 3.5 Pro is currently testing with partners and will be broadly available “as soon as it’s ready.” The company is already pre-training Gemini 4.
Why is the cybersecurity model restricted? Google cites the “dual-use nature” of code vulnerability finding — the same capability that helps defenders can help attackers. The restricted distribution is a safety measure, but also a competitive moat.
🔍 THE BOTTOM LINE
The model war’s next phase is not about who has the smartest AI. It is about who can run agents cheaply enough that businesses deploy them by the thousand, not the dozen. Google’s three-model release — efficient workhorse, fast throughput, restricted specialist — covers the three bases that matter. The cybersecurity model is the tell: vertical-specific AI is where the differentiation lives, and Google just claimed the first one.
