Google is turning Gemini’s “Flash” line into the backbone of its AI strategy, and the new 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber drops are clearly designed for a world where AI agents are no longer demos, but production infrastructure.

TL;DR: 3.6 Flash is the new everyday workhorse, 3.5 Flash-Lite is the speed-obsessed sidekick for high-volume tasks, and 3.5 Flash Cyber is Google’s specialized defender for software security.

If 3.5 Flash was the proof of concept, 3.6 Flash is the “default mode” future

Google is very openly positioning Gemini 3.6 Flash as the new baseline model for everyday AI work – coding, research, document parsing, and multimodal tasks. It’s built directly on developer and customer feedback from 3.5 Flash, which launched earlier this year, and the theme of this update is straightforward: do more, with less, and charge less for it.

Under the hood, 3.6 Flash is still a multimodal model with a 1 million token context window and support for text, images, audio, video, and PDFs, but Google has tuned it aggressively for token efficiency. On the Artificial Analysis Index, the company says 3.6 Flash uses 17% fewer output tokens than 3.5 Flash for comparable tasks, and in some coding benchmarks like DeepSWE, the reduction can go as high as 65%. That matters because in the AI world, tokens are money: fewer tokens for the same job means lower bills for developers and enterprises running large agent workflows.

Pricing reflects that angle. 3.6 Flash comes in at $1.50 per million input tokens and $7.50 per million output tokens, a cut from 3.5 Flash’s output pricing. For teams building large fleets of AI agents, that’s the difference between “interesting prototype” and “this can actually sit in production without the CFO panicking.”

Beyond efficiency, Google is leaning on performance benchmarks to make the case that 3.6 Flash isn’t just cheaper, it’s actually better. On DeepSWE, a benchmark for software engineering tasks, 3.6 Flash hits 49% versus 37% for 3.5 Flash, while on MLE Bench for machine learning research, it jumps from 49.7% to 63.9%. On OSWorld-Verified, a benchmark for computer use – think controlling apps and systems as an agent – it improves from 78.4% to 83%.

That last part is important, because Google is now treating “computer use” as a built-in client-side tool in the Gemini API and Gemini Enterprise, turning 3.6 Flash into an agent that can interact with screens, apps, and interfaces rather than just text. Google’s own case studies showcase 3.6 Flash using managed agents to parse financial data and earnings transcripts, orchestrating multi-agent code migrations, or helping build visual tools like texture extractors and interactive theme studios.

All of this is wrapped in a stronger safety story. 3.6 Flash ships with enhanced “Frontier Safety” controls for sensitive areas like chemical, biological, radiological, nuclear, and cyber offense misuse, with Google claiming it is substantially more resistant to jailbreak attempts while still reducing needless refusals for legitimate use. For regulators and large enterprises, that safety positioning is no longer optional – it’s part of how you sell an AI platform.

In terms of availability, Google isn’t treating 3.6 Flash as an experimental model. It’s rolling out broadly to developers via the Gemini API in Google AI Studio and Android Studio, to enterprises via the Gemini Enterprise Agent Platform and Gemini Enterprise app, and to consumers in the Gemini app itself. In other words, if you open Gemini, there’s a good chance 3.6 Flash is already doing the work behind the scenes.

Flash-Lite grows up: 3.5 Flash-Lite is the high-speed specialist

If 3.6 Flash is the generalist, 3.5 Flash-Lite is the specialist that lives where latency and cost are non-negotiable. Google describes it as “the fastest, most cost-effective model in the 3.5 series,” built for high-throughput scenarios like agentic search, large-scale document processing, and subagent workloads that don’t always need deep reasoning.

On paper, Flash-Lite looks like a classic “small but sharp” tool. It runs at around 350 output tokens per second according to Artificial Analysis – faster than previous Flash-Lite generations and even faster than 3.5 Flash itself. Pricing is notably aggressive: $0.30 per million input tokens and $2.50 per million output tokens, a fraction of what you’d pay for a larger model. For teams building AI agents that need to respond in near real-time, or process thousands of items per minute – think e-commerce metadata extraction, receipt summarization at scale, or running multiple design variations – this is the kind of model you park behind the scenes and let it churn.

Despite its “Lite” branding, 3.5 Flash-Lite still supports a full multimodal input stack: text, images, video, audio, and PDFs, though its outputs are text-only. It keeps the same 1 million token context window and up to 65,536 output tokens, which means it’s more than capable of handling long documents or multi-step workflows when needed.

In benchmarks, it often punches above its weight. On Terminal-Bench 2.1, which focuses on coding and agentic tasks, 3.5 Flash-Lite reaches 54% versus 31% for 3.1 Flash-Lite. On GDM-MRCR v2 for long-context understanding, it climbs from 60.1% to 72.2%, and on GDPval-AA v2, a real-world task execution benchmark, it jumps from 642 to 1140. Perhaps more surprising, Google says Flash-Lite is now outperforming Gemini 3 Flash on several agentic and coding metrics, including SWE-Bench Pro (54.2% versus 49.6%) and OSWorld-Verified (74% versus 65.1%).

One subtle but telling design choice: 3.5 Flash-Lite defaults to “minimal thinking” for classification and extraction tasks, optimizing for speed and cost, but can be configured with higher “thinking levels” for multi-step subagent workflows. That’s a clear nod to how developers are actually using these systems – not as single monolithic models, but as networks of agents with different roles, some needing depth and others simply needing to move fast.

Just like 3.6 Flash, Flash-Lite also supports computer use as a built-in tool, which makes it more competent at tasks like navigating interfaces or executing workflows across surfaces. Early customers Google cites – including names like Ashler, Palo Alto Networks, and Ramp – are focusing on its sweet spot: speed plus intelligence plus cost efficiency for scaling agentic workflows.

In practice, you can imagine a pattern where 3.6 Flash acts as a “master” agent, orchestrating tasks, while Flash-Lite runs as a fleet of subagents doing heavy lifting on receipts, product data, or various design explorations. Google even highlights an example where 3.5 Flash-Lite instantly generates 25 unique web design concepts, working alongside 3.6 Flash.

Gemini 3.5 Flash Cyber: AI moves deeper into code security

The third model in this launch, Gemini 3.5 Flash Cyber, feels different in tone from the other two. This isn’t a general-purpose tool heading straight into the Gemini app. It’s a specialized, lightweight cybersecurity model, built on top of 3.5 Flash and fine-tuned to find, validate, and patch vulnerabilities at scale.

Google has been investing in automated security systems for years – with internal tooling like CodeMender, its code security agent that hunts down and helps fix critical vulnerabilities across large codebases. 3.5 Flash Cyber is essentially built to push that idea further, using the Flash architecture to detect and remediate issues more efficiently than larger general models, and at a lower price per token.

In benchmarks, Google points to CyberGym, a popular evaluation for cybersecurity tasks, where multiple 3.5 Flash Cyber agents within CodeMender collaborate to produce a unified report and reach competitive frontier-level performance. The idea here isn’t just “AI spots bugs”; it’s “AI agents work together on security like a team,” which aligns with Google’s broader “agentic” narrative for Gemini.

But because this is a dual-use technology – powerful cybersecurity tooling can be a defender’s dream and an attacker’s weapon – Google is deliberately holding Flash Cyber back from the usual Gemini rollout. Instead, it’s launching as part of a limited-access pilot, exclusively available to governments and selected “trusted testers” through CodeMender. The pitch is that frontline defenders get a head start in finding and fixing critical vulnerabilities before they’re exploited, while broader misuse is mitigated by restricted access.

For the wider developer community, this might feel a bit out of reach for now, but strategically, it signals something important: major AI platforms are no longer just talking about security as “how we keep our models safe.” They are actively building models whose main job is keeping your code safe.

The bigger picture: agents, efficiency, and the road to Gemini 4

One of the more interesting threads running through Google’s announcement is how openly it talks about AI agents instead of just models. The Flash line is described as hitting the “sweet spot of efficiency and quality to enable scaling agentic workflows,” and each new variant is framed in terms of how it fits into those workflows rather than just its standalone capabilities.

3.6 Flash is the “workhorse” – the default choice for orchestrating complex tasks, handling multimodal input, and driving knowledge work. 3.5 Flash-Lite is the scaler – tuned for high throughput, low latency, and minimal cost, ideal for subagents that do repetitive but essential jobs. 3.5 Flash Cyber is the specialist – dropped into security pipelines to cooperatively analyze code and patch vulnerabilities.

Taken together, they paint a picture of a future where AI systems look less like singular chatbots and more like distributed teams, with different models filling distinct roles. That’s a big shift from the early Gemini narrative, which was mostly about single flagship models, and it helps explain why Google is making such a fuss about token efficiency, pricing, and computer use capabilities.

Google also uses this launch to quietly update the roadmap. It confirms that Gemini 3.5 Pro is still in testing with partners, with plans to make it broadly available once ready, and says the team has already started its most ambitious pre-training run yet for Gemini 4. The subtext: while developers are adapting to 3.6 Flash and Flash-Lite, Google’s next generation is already in motion.

For users and builders in the US, especially those thinking about real-world deployment – from SaaS tools to enterprise workflows – this is the kind of release that matters more than a flashy keynote. It changes the default cost structure, improves safety guarantees, and gives you more granular tools to assemble AI agents that feel like practical, sustainable infrastructure rather than experimental toys.