Artificial Intelligence

Google Shrinks Gemini Down to a Cheap, Fast, Specialized Family

Instead of one big flagship, Google now sells the Flash line in three sharply different sizes — a fast generalist, a tiny high-throughput sub-agent, and a cybersecurity specialist — all priced to make AI calls economically viable at scale.

• Gemini 3.6 Flash is the new general-purpose workhorse, improving on the coding and finance benchmark results of its predecessor.
• Gemini 3.5 Flash-Lite is built for high-throughput, low-latency jobs such as agentic search, document parsing and sub-agent execution.
• Gemini 3.5 Flash Cyber is a purpose-built security model, aimed at detecting and explaining cyber threats.

The move marks a shift in how frontier labs sell intelligence. For years, the headline story has been bigger models with higher raw scores on leaderboards. Google is now doing the opposite with its Flash family: keeping the same "flash" branding while carving the capability curve into distinct products, each optimized for a specific kind of work rather than competing for a single top score.

The economics are the real story. 3.5 Flash-Lite is priced at roughly 30 cents per million input tokens and 2.50 dollars per million output tokens. At those numbers, a service can run dozens of AI calls per user interaction and still hold a margin — something that was borderline uneconomical a year ago. The model is also notably concise in its output, which keeps the pricier output-token side of the bill small. That combination of speed, brevity and price is designed for agentic architectures, where a single user request spawns a chain of internal tool calls, each handled by a cheap fast model before the final answer is assembled.

The split also changes what "specialist" means in a model family. Flash Cyber exists so security analysts can feed logs, alerts and threat context into a model tuned and red-teamed for that domain, rather than relying on a general assistant to improvise. That is a more defensible stance than prompting a flagship model to "act like a SOC analyst" — the specialist can be validated, versioned and restricted on its own.

Together the three releases frame the practical future of frontier AI: not one smarter oracle, but a toolkit where a generalist, a high-volume worker and a domain specialist each do the job they are cheapest to run at. The race is no longer just about who has the highest leaderboard number, but who can make the whole stack of calls cost enough to actually ship.

Knowledge takeaway: Gemini 3.6 Flash is the updated general-purpose workhorse; 3.5 Flash-Lite is a high-throughput, low-latency sub-agent model priced near $0.30/$2.50 per million input/output tokens; 3.5 Flash Cyber is a purpose-built cybersecurity specialist.