Gemini 1.0 felt like Google changing the shape of the model race.
three sizes, one launch
Yes, it was a direct answer to GPT-4. That part was obvious.
The more interesting move was the product shape: Ultra, Pro, and Nano; multimodal from the start; and a launch that made placement part of the conversation instead of just raw benchmark ranking.
That mattered because it forced a more honest question. Where should intelligence live. On the device. In the product flow. In the expensive frontier route. The answer was not one model for everything, and Gemini said so out loud.
I remember the launch mostly for the way it changed the default architecture sketch:
small on-device model -> quick local work
pro model -> normal multimodal product flow
ultra model -> hard reasoning or expensive review
That is a real shift. Once you have that ladder, the product stops pretending there is one correct route for every task.
The other thing Gemini 1.0 did was make mixed input feel less like an exception. A screenshot, a voice note, a chart, and a document all belonged in the same conversation. That sounds obvious now. It did not feel obvious before.
remove the glue first
If I were shipping against that model family, I would strip out the glue code before adding more prompt text. For a billing flow, the request could look more like this:
user uploads a screenshot of a charge
user adds: "this is not mine"
client keeps the image, visible text, and account context together
pro model handles the first answer
ultra only gets the case if the first pass cannot tell whether this is fraud or a duplicate charge
That is the point. The model size tells you where the work should sit, but the product still has to keep the important inputs intact long enough for the model to see them.
If I were writing the brief, I would have to say which route answers on-device, which route stays in the app, and which route is worth the expensive call. That is more annoying to write and much easier to operate.
Gemini 1.0 made that trade visible early, and once you notice it you stop designing as if all intelligence belongs in one place.
Related posts

About Jeremy London
Engineering leader and builder in Denver. I write about AI platforms, agents, security, reliability, homelab infrastructure, and the parts of engineering work that have to survive production.