AI Pulse · AI Trends Pulse

Google released three new Gemini models yesterday, and the story is not what they can do. It’s what they cost when you run them thousands of times a day.

The headline model, 3.6 Flash, uses 17% fewer output tokens than its predecessor for the same task, according to Google’s announcement. That matters because most enterprise AI spend now comes from agents, not one-off queries. An agent that checks inventory, routes a support ticket, or summarises a call might hit the model fifty times in a single workflow. Seventeen percent fewer tokens per call compounds fast. It’s a direct play for volume buyers who care more about the monthly bill than leaderboard rankings.

The other two models follow the same logic. Flash-Lite is stripped down for tasks that don’t need reasoning depth. Flash Cyber has been security-vetted for regulated environments. Neither is trying to beat GPT-4 on a benchmark. Both are trying to slot into repeating, high-frequency workflows where speed and cost predictability matter more than raw capability.

This is the shift worth watching. The race is no longer just about who builds the smartest model. It’s about who builds the cheapest one that’s still good enough to automate the boring, repetitive work that actually runs a business. If you’re planning to deploy agents at scale, token efficiency is now as important as accuracy. The models that win in 2025 will be the ones that let you run a hundred automations for the price you used to pay for ten. That’s the kind of cost structure we bake into tools like the Omni Command Centre, where every agent call adds up and efficiency is the difference between a pilot and a profit centre.