Last Updated: 23 July 2026
Google just rolled out three new Gemini models on July 21: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber. The first two are already generally available, so they're stable and ready to use, not just early previews. Flash Cyber is the odd one out. It's only available through CodeMender, Google's code security agent, and it isn't open to the public at all. As for Gemini 3.5 Pro, the more advanced model many people were expecting, it's still being tested with partners, and Google hasn't given any indication of when that will change.
That leaves an obvious question for anyone keeping tabs on Gemini: which of these models actually applies to you, and what does it change in practice? Google claims Gemini 3.6 Flash needs less output tokens than Gemini 3.5 Flash, and that Flash-Lite can crank out 350 output tokens every second. Both numbers are really pointing at the same thing: this release is about making AI agents run faster and cheaper at scale, not about launching yet another chatbot.
To make sense of it all, it helps to untangle a few things that tend to get lumped together: what each model is actually built for, where you can get your hands on it, and what it costs through the developer API compared to a regular Gemini subscription.
Among the new releases, Gemini 3.6 Flash covers the widest range of tasks. Google positions it as its workhorse model for coding, multimodal tasks, general knowledge work and AI-agent workflows that involve several steps. Gemini 3.5 Flash-Lite is built for a different kind of job: high-volume work where speed and lower running costs matter more than handling the most complex tasks, such as document processing or agentic search. Gemini 3.5 Flash Cyber is a specialised security model, fine-tuned to work with CodeMender, and it comes with restricted access rather than a public rollout.
It helps to think about these three releases by workload rather than by ranking them from best to worst. Gemini 3.6 Flash suits broader, more complex agent work. Flash-Lite is aimed at large volumes of faster, repeated tasks, which makes it a different tool rather than a lesser one. Flash Cyber is not part of the Gemini app or the standard public API, so it will not appear as an option for most developers or everyday users.
It is also worth remembering that intended use and public availability are two separate things. A model can be described as ideal for a certain job while still being unavailable to most people, which is exactly the situation with Flash Cyber. Gemini 3.5 Pro, meanwhile, sits outside this release entirely. It is not one of the three new models, and Google has not tied it to this announcement in any concrete way.
As of now, you can get to both Gemini 3.6 Flash and Gemini 3.5 Flash-Lite through the Gemini API, Google AI Studio and Android Studio. Gemini 3.6 Flash also lives in Google Antigravity, the Gemini Enterprise Agent Platform and the Gemini Enterprise app. Flash-Lite is available in many of the same places, including the Gemini Enterprise Agent Platform and the Gemini app, and Google says it's gradually reaching Google Search too.
That "gradually" is doing a lot of work in that sentence. It doesn't mean Flash-Lite is instantly available to everyone searching on Google. There's no published country-by-country schedule, so some users will see it show up before others. Enterprise customers should also pay attention here, since an admin has to manually turn on the Gemini 3.6 Flash feature before anyone on their team can use it, something that's easy to miss.
It helps to keep developer access mentally separate from everyday consumer access. Building with a model through the API is a different path than simply seeing it appear inside the Gemini app, and enterprise rollout moves on its own timeline entirely. Flash Cyber sits outside all of this. It isn't part of the general rollout at all and remains confined to its own limited pilot.
Google says Gemini 3.6 Flash gets by using 17% fewer output tokens than Gemini 3.5 Flash, a figure based on the Artificial Analysis Index. Output tokens are basically the text, code or other material a model spits out in response to a request, and since API pricing depends partly on how much a model generates, using fewer tokens can bring the cost of a task down, though how much depends on the workload.
There's more to it than efficiency, though. Google also says the model needs fewer reasoning steps and tool calls to get through multi-step work, which is a big deal for AI agents that have to plan, verify information and take action across several stages instead of just answering one question. On top of that, 3.6 Flash comes with Computer Use built in through the Gemini API and Gemini Enterprise, letting it interact with software directly while working through a task.
Picture a coding agent that needs to check a file, use a tool, and rework its output based on what it finds, doing that loop over and over until the job is done. That's essentially the workflow Google built 3.6 Flash for. Since these efficiency and performance numbers come straight from Google, they're best treated as the company's own benchmark claims rather than something independently verified, even if they still capture a meaningful part of what this release is about.
Gemini 3.6 Flash claims a 17% cut in output tokens versus its predecessor, fewer reasoning steps for agent tasks, and built-in Computer Use, aimed squarely at faster, cheaper multi-step AI workflows.
Speed and volume are what Gemini 3.5 Flash-Lite is really about, not tackling the most complex problems out there. Google says it can produce 350 output tokens every second, positioning it as the fastest, most budget-friendly model in the entire 3.5 Flash family. Put simply, high-volume work covers things like processing a flood of requests, doing the same task repeatedly, or working through big stacks of documents, rather than one long, nuanced conversation.
Out of the box, Flash-Lite runs at a minimal thinking level, which Google says keeps things fast and cheap for lighter tasks like classification, routing, or pulling structured data from raw input. When the job gets harder, though, like autonomous subagents writing code, executing terminal commands, or juggling external tools across several steps, Google recommends stepping up to medium or high thinking instead. Keeping it on minimal for those multi-step jobs risks tools quitting early, before the work is actually finished.
Flash-Lite also mirrors Gemini 3.6 Flash: it includes Computer Use as a built-in tool and supports a large context window. One practical note for anyone migrating an older system: custom temperature, top-K and top-P settings won't work anymore, and custom frequency or presence penalties will trigger an API error. In the end, it's speed that separates Flash-Lite from the broader 3.6 Flash release, not overall capability.
Built for throughput over depth, Flash-Lite delivers 350 tokens per second, flexible thinking levels for simple or complex tasks, and native Computer Use, all tuned for fast, high-volume agent work.
Google prices Gemini 3.6 Flash at $1.50 per million input tokens and $7.50 per million output tokens. Gemini 3.5 Flash-Lite comes in a lot cheaper, at $0.30 per million input tokens and $2.50 per million output tokens. Input tokens are basically whatever you send the model, a prompt or a document, while output tokens are what it hands back to you.
Keep in mind these are developer API rates, not the price you'd pay for a regular Gemini subscription. So they're really meant for businesses and developers building products, not someone just chatting with the Gemini app on their phone. Google also points out that 3.6 Flash's output pricing beats the $9 per million output tokens that the earlier Gemini 3.5 Flash charged, even though input pricing has held steady at $1.50 per million tokens.
That price difference lines up with the split between the two models. Gemini 3.6 Flash is the broader, more capable agent model, and its cost reflects that. Flash-Lite's lower rate makes more sense for high-volume work where requests pile up fast. Either way, what you actually pay depends on how much you're sending and generating, so there's no universal answer for which model is cheaper.
Regular Gemini users and most developers cannot access Gemini 3.5 Flash Cyber. Google says the model will be available through a limited-access pilot for governments and trusted partners via CodeMender, rather than through the Gemini app or the public API that the other two models use.
Flash Cyber is built on Gemini 3.5 Flash and fine-tuned specifically to find, validate and patch software vulnerabilities. Multiple Flash Cyber agents can work together within CodeMender, each examining different code paths, before their findings are combined into a single report. Google describes the restricted rollout as intentional, saying cybersecurity tools carry dual-use risks and that limiting early access helps give defenders a head start while reducing the chance of broader misuse.
Google has pointed to internal testing to support the model's usefulness, including results from its own CyberGym evaluation, but these figures come from Google rather than independent researchers and should be read with that in mind. No public self-service access or API pricing has been announced for Flash Cyber, so for now it remains a pilot program rather than a model most readers will be able to try.
Gemini 3.5 Pro is still testing with partners, and Google says it will become broadly available once it is ready, without offering a specific public release date. That leaves the more advanced model absent from this announcement entirely. All three of the models released on July 21 are Flash variants, built around efficiency and speed rather than the expanded capability typically associated with a Pro release.
Google has said it has begun what it calls its most ambitious pre-training run yet, aimed at Gemini 4, which suggests its attention is also turning toward what comes after the current Pro model. Reuters has reported that Gemini 3.5 Pro was originally expected around June,
Taken together, the emphasis on lower output-token use, high generation speeds and reduced API rates in this release may suggest that Google is prioritising models that cut the cost and time needed to run AI agents, even as its more capable Gemini 3.5 Pro model continues testing behind the scenes. This is an interpretation, not a stated reason from Google, and it should not be read as confirmation that the Flash releases were driven by the Pro delay.
Gemini 3.6 Flash is Google's new general-purpose Flash model built for coding, knowledge work, multimodal tasks and multi-step AI agents. Google says it uses 17% fewer output tokens than Gemini 3.5 Flash and needs fewer reasoning steps and tool calls in certain workflows. It's the broader option in this new lineup for developers, while Gemini 3.5 Flash-Lite is meant for faster, high-volume tasks instead. Both models come with built-in tools, including Computer Use, and are available through Google's public Gemini developer platforms.
Gemini 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens on the Gemini API, and Gemini 3.5 Flash-Lite costs $0.30 per million input tokens and $2.50 per million output tokens. Input tokens cover material sent into the model, while output tokens cover what comes back out. These are developer API rates rather than consumer subscription prices. Actual costs depend on how much you send, how much the model generates in response, and the service tier being used.
Gemini 3.5 Flash-Lite is Google's fastest, lowest-cost model in the 3.5 line, built specifically for high-throughput execution. It's suited to work like document extraction, classification, routing, structured data processing and simpler agent tasks. Google puts its output speed at 350 tokens per second. Minimal thinking is the default setting, aimed at simple, latency-sensitive requests. For autonomous subagents that rely on tool calls, code execution or multi-step reasoning, though, Google advises using medium or high thinking levels to keep tools from ending early on longer, more complex work.
No, Gemini 3.5 Flash Cyber isn't part of a typical public rollout. Google intends to give governments and trusted partners first access through a limited CodeMender pilot. The model is built on Gemini 3.5 Flash and fine-tuned specifically to detect, validate and patch software vulnerabilities. According to Google, this restricted approach accounts for the dual-use risks that come with powerful cybersecurity tools. Since there's no public self-service access or published API pricing yet, most Gemini users have no way to try Flash Cyber through standard Google tools.
According to Google, Gemini 3.5 Pro remains in partner testing and will become widely available once it meets that bar. No public launch date, model card or pricing has been shared for Pro as part of the July 21 announcement. Gemini 3.6 Flash and Flash-Lite are distinct releases and aren't stand-ins for the Pro model still on the way. Reuters reported that Pro was originally expected in June, but Google itself has only confirmed continued partner testing with a broad release to follow eventually.