Google launches Gemini 3.6 Flash and 3.5 Flash-Lite – Gemini 3.5 Pro still missing

What this is about

In a July 21 blog post, Google introduced two new models in its cheap, fast "Flash" line: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. According to Google, both are designed to run AI agents at scale and are meant to deliver "efficiency, low latency and reliability." Also new is Gemini 3.5 Flash Cyber, a specialized cybersecurity model. Mashable covered the news. The big model, Gemini 3.5 Pro, is still missing.

The three models at a glance

Gemini 3.6 Flash is said to improve on 3.5 Flash in coding, knowledge work and multimodal tasks (text, images and more), and it's cheaper anyway. Google published benchmarks showing gains in several areas; Mashable doesn't give the individual figures. The model also has stronger safeguards against jailbreaks (tricks that bypass safety rules) related to chemical, biological, radiological and nuclear misuse as well as cyberattacks, while avoiding unnecessary refusals of legitimate requests.

Gemini 3.5 Flash-Lite is built for high throughput and short response times, for example in agentic search and document processing. According to Artificial Analysis figures cited by Google, it manages 350 output tokens per second. In agent benchmarks, it's said to beat the previous Flash-Lite generation.

Gemini 3.5 Flash Cyber is used in CodeMender, Google's code security agent. Because of the dual-use risk – a model that finds vulnerabilities can also help attackers – access stays limited: Google plans a pilot program for governments and selected partners only.

What about Gemini 3.5 Pro?

At I/O in May, Sundar Pichai announced that Gemini 3.5 Pro would launch in June. Instead, Google now says the model is being "tested with partners" and will arrive "when it's ready." At the same time, training runs for a future Gemini 4 have reportedly begun. Mashable points to pressure from competitors: Anthropic and OpenAI have released new flagship models, Fable 5 and GPT-5.6 Sol, and the Chinese lab Moonshot has launched Kimi K3, a cheaper but competitive open-source model.

What has happened since

Other reports from the following months (CNET, Ars Technica, note.com) show how things developed. They don't directly concern the Flash models, but they put the situation in context. Gemini 4 Argon arrived in late September, which Ars Technica calls Google's first new flagship model in 11 months. About two weeks later, on October 28 (US time), came Gemini 4 Carbon, which Google calls its most powerful model yet, with a focus on cybersecurity. According to Google, it ranks first in Artificial Analysis's overall index with 59 points. Ars Technica notes that Carbon beats Argon only narrowly on every test and that the gap to competitors (GPT-6 Astra, Claude Opus 5.5) is just a few points. It could also be a post-training upgrade of Argon.

So far, Carbon is available only to selected cyber defenders in the "Fairwind Program," with no date for everyone else. note.com puts API pricing at $2/$10 per million tokens at launch and $4/$20 later. The safety scores (such as 0% fraud in tests) are vendor claims, measured without safeguards in place. According to CNET, Google also introduced the world model Genie 4, which can handle audio for the first time. There's a waitlist.

Our take

The Flash models matter mostly to developers and companies that want to run agents cheaply and fast. The lower price of 3.6 Flash and the high speed of Flash-Lite are the tangible gains. All performance claims come from Google itself; the sources mention independent testing only for speed (Artificial Analysis). One pattern stands out: the strongest models – Flash Cyber, and later Argon and Carbon – are initially available only in limited form, justified by misuse risks. For regular users, that means the Flash variants are, for now, the main ones you can actually try. It remains to be seen when the top models will become widely available.