Before you spend on hardware, run the exact open model you'd host at home on your real work, from one rented account, and step up sizes in minutes

Don't buy a box for AI until you've rented the model you plan to run on it. Yesterday's Lesson sized up what your old hardware can handle. Today's move comes first: run that exact model on your real work, from a browser, before anything lands on your desk.

Most people who want AI running locally haven't answered the one question that decides the purchase. Is the small model good enough for what I actually do? Hardware answers that the slow way. OpenRouter answers it this afternoon. It puts hundreds of models behind one account and one API, open-weight and frontier alike, and you pay as you go instead of up front. The open models you'd download for your own machine are on the shelf. Google's Gemma 4 family, for example, includes a 26B mixture-of-experts model that activates about 4 billion parameters per token, listed from roughly four cents per million input tokens, and a 31B dense model one step up. Google's own sizing puts the 26B at about 14GB of memory at 4-bit, which is within reach of a desktop with 24GB or more. OpenRouter isn't the only door, either. Groq runs open models very fast on chips it builds itself, and Fireworks AI hosts open models and adds fine-tuning when you want one shaped to your own work.

So run the test. Pick one task you do every week: summarizing a contract, drafting replies to customer email, turning meeting notes into action items. Open OpenRouter's chat, choose the smallest model that fits the machine you were about to buy, and feed it last week's real inputs. Then step up one size and run the same inputs again. That's minutes, not a procurement cycle.

Both outcomes are good ones. If the small model handles the job, you know exactly what hardware you need, no more and no less, and you can buy with confidence. If it doesn't, you found out before you owned a box that can't run the model your work actually requires. Either way, you stop guessing.

And if the honest answer is that you don't have the hardware, the admin, or the time to own your IT, stop right there and keep renting. That's a strategy, not a compromise. I'd much rather see a team running a good model through a rented account this week than a new workstation sitting in its box until spring.

One setting before you paste in anything sensitive. OpenRouter says it does zero logging of your prompts unless you opt in, but the providers it routes to set their own policies. Set the data_collection option to deny and requests go only to providers that don't collect your data. You can make that the default for your whole account in your privacy settings. Flip it first. It's the same habit as the training toggles we covered two weeks ago, one layer down.