Google launches Gemini 3.6 Flash & Flash-Lite; confirms Gemini 4 is in training

Google just had a busy Tuesday. On July 21, 2026, the company shipped not one, not two, but three new Gemini models—Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a security-focused variant called Gemini 3.5 Flash Cyber—while quietly dropping one of the more significant lines of the year near the bottom of its announcement: Google has already begun pre-training Gemini 4.

It’s a release that says a lot about where the AI industry’s attention currently sits: not on chasing the next flagship headline model, but on making the middle of the model stack faster, cheaper, and more efficient, all while the next generational leap quietly gets underway in the background.

What’s Actually New

Gemini 3.6 Flash: The Workhorse Gets an Upgrade

Described by Google as its workhorse model, Gemini 3.6 Flash follows up on Gemini 3.5 Flash, which launched roughly two months earlier at I/O 2026. The headline improvement isn’t raw capability—it’s efficiency. According to the Artificial Analysis Index, 3.6 Flash uses 17% fewer output tokens than its predecessor while completing multi-step tasks in fewer reasoning steps and tool calls.

That efficiency translates directly into cost. Gemini 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens—down from $9 per million output tokens for 3.5 Flash. For coding specifically, Google says the model produces higher-precision output with fewer unwanted edits and less looping during execution, pointing to gains on the DeepSWE benchmark as evidence of more reliable, production-ready code generation.

Gemini 3.5 Flash-Lite: Built for Speed and Scale

Alongside 3.6 Flash, Google introduced Gemini 3.5 Flash-Lite, its fastest and most cost-effective model in the 3.5 line. Running at roughly 350 output tokens per second, Flash-Lite is aimed squarely at latency-sensitive, high-throughput workloads like agentic search and document processing. It’s priced at $0.30 per million input tokens and $2.50 per million output tokens, undercutting most competitors at a similar price point, while still outperforming earlier Flash-Lite generations on agentic tasks.

Gemini 3.5 Flash Cyber: A Model Google Isn’t Opening Up

The most unusual release of the day is Gemini 3.5 Flash Cyber, a specialized variant fine-tuned for vulnerability detection and patching. Rather than releasing it broadly, Google is pairing it with its CodeMender security agent and limiting access to a pilot program for governments and trusted partners. Early enterprise testers reportedly include Salesforce, Robinhood, and Palo Alto Networks.

The gating here is notable. Google has generally leaned toward broad availability with its models, so a deliberate decision to restrict an offense-capable security tool signals real caution about how this kind of model could be misused if released as a self-serve API.

Where’s Gemini 3.5 Pro?

Conspicuously absent from Tuesday’s releases is Gemini 3.5 Pro, the flagship model many developers have been waiting on. Google confirmed it’s currently testing with partners and reiterated plans to make it broadly available once it’s ready, without offering a firm date. For now, the Flash line is doing the heavy lifting while Pro continues its testing phase.

The Bigger Story: Gemini 4 Pre-Training Has Begun

Buried near the end of Google’s announcement is the line generating the most attention: Google says its DeepMind team has started its most ambitious pre-training run yet, for the next generation of models—Gemini 4.

That’s a meaningfully different claim than a typical roadmap tease. Pre-training is the earliest, most resource-intensive phase of building a large model, and confirming it has already begun suggests Google is treating Gemini 4 as a genuine step-change rather than an incremental release. It’s also worth being precise about what was and wasn’t said: this is pre-training that has started, not a release date or capability preview. Read it as an early signal of ambition, not a delivered product.

Built-In Safety Measures

Alongside the efficiency gains, Google says Gemini 3.6 Flash ships with enhanced Frontier Safety safeguards specifically targeting chemical, biological, radiological, and nuclear misuse, as well as cyber offense risks. The company states these safeguards make the model substantially more resistant to jailbreak attempts, while simultaneously being tuned to avoid over-refusing legitimate, beneficial requests—an attempt to strike a balance between safety and usability that AI labs have struggled with as models get both more capable and more scrutinized.

Why This Release Matters

This wasn’t a splashy flagship launch, and that’s arguably the point. A few things about Tuesday’s announcement stand out:

• The AI price war keeps intensifying. Cheaper, faster Flash-tier models suggest competitive pressure across the industry is pushing costs down even as capability climbs.
• Efficiency is becoming its own competitive axis. A 17% cut in token usage isn’t a flashy benchmark score, but it compounds significantly at enterprise scale, where cost per task often matters more than raw intelligence.
• Security-specific models are becoming their own category. Gemini 3.5 Flash Cyber suggests major labs are increasingly building specialized, access-restricted models for high-stakes domains rather than relying on general-purpose models with guardrails bolted on.
• The roadmap is already looking past this release. With Gemini 3.5 Pro still in partner testing and Gemini 4 pre-training already underway, Google is signaling that today’s Flash updates are a bridge, not a destination.

Conclusion

Google’s Tuesday releases were less about a single headline model and more about tuning the middle of its Gemini lineup for cost, speed, and specialized use cases like cybersecurity—while a much bigger swing, Gemini 4, is already quietly taking shape in pre-training. For developers and enterprises, that means cheaper and faster models are available right now. For anyone watching the broader AI race, the real story might be the one sentence Google tucked near the end of its post: its most ambitious training run yet is already underway.