Every AI coding tool you can buy has the same cost structure: each completion, each agent step, each file read is an API call somebody pays for. That is why the free tiers throttle, why the plans are per seat, and why the pricing page always has a usage table.
Osmium has that cost too, for cloud models. But the interesting case is the other one. When the model runs on your GPU through Ollama or LM Studio, we are not in the loop at all. There is no request to our servers, no token to count, and no bill arriving at the end of the month. Charging for it would mean inventing a cost.
So the free tier is the actual product
Not a trial. Not a limited mode with the good parts greyed out. The editor, the terminal, git, the debugger, search, MCP, and the Agent, Plan, Ask, Analyze, and Debug modes all work with a local model, permanently, with no card.
What money buys is frontier cloud models with the keys handled for you, and the heavier modes — Architect, Arena, and Fleet. That is a real upsell for real work. It is not a tax on the basics.
The honest catch
A 7B model on your laptop is not Claude. On a hard multi-file problem the frontier models are meaningfully better, and pretending otherwise would waste your time. What we have found is that a large share of daily work — tests, small refactors, reading unfamiliar code, boilerplate — is handled fine by a local coder model, and that share is free.
The right way to find out where your line is: install it, point it at your own repository, and use Arena to race a local model against a cloud one on your actual code. Then you are deciding with evidence instead of a benchmark someone else ran.
The part we did not expect
The privacy argument turned out to matter more than the price. Teams that cannot paste source into a third-party service were not choosing between AI tools — they had no AI tooling at all. A local model does not need a policy exception, because there is no request to review.