Choosing models and controlling cost
GNexusOS does not resell AI. You bring your own provider keys and you pay that provider directly, at their prices, with no margin added.
That is good for you and it has one consequence worth understanding: the app cannot tell you what you have spent, and does not pretend to.
Two separate bills
What you pay for GNexusOS is set out on the pricing section of the homepage and in the Terms of Service. It is not metered by usage.
What you pay your AI providers is between you and them. It depends entirely on which models you pick and how much work you give them.
Some capabilities need their own separate account on top: web search needs a search provider's key, and voice needs a speech provider's key. See Connections.
Why the dashboard sends you elsewhere
The provider card on the dashboard says only:
Usage data available in provider console.
and gives you a button to open it.
This looks like a missing feature. It is a deliberate one. The only authoritative record of your spend is the provider's own billing page, and a figure estimated locally would be wrong in the one domain where being wrong is worst. A number you cannot trust is worse than a link to the number you can.
Open that console once a week for the first month. It is the only way to build an instinct for what your usage actually costs.
Picking a model for the job
The model dropdown sits next to the provider in the dispatch bar. Your choice is remembered for that provider — it is not a one-off for this task.
| Kind of work | What to reach for |
|---|---|
| Sorting, tagging, extracting fields | The smallest model that works |
| Summarising something you will read yourself | Small |
| Drafting you will edit anyway | Mid |
| Writing that goes out as-is | The best you have |
| Reasoning, planning, judgement calls | The best you have |
| Anything that must be right first time | The best you have |
The rule behind the table: pay for judgement, not for typing. A large model re-reading a spreadsheet to pull out three numbers is the most common way to spend money for nothing.
Running locally costs nothing per call
A local provider runs models on your own machine, so there is no per-call charge and no key to enter. It is slower and generally less capable, which makes it a good fit for exactly the high-volume, low-judgement work in the top rows of that table.
If you connect one, remember it takes no API key — a key field left blank is correct, not an error. See What the error messages mean.
Settings that spend money quietly
None of these are traps, but each one is a call you might not have counted.
The prompt wand (✦) rewrites your instruction before sending, which is an extra call every time you use it. Worth it on a hard request, wasteful on "summarise this".
Computer use has its own model, set in System Config → Control and independent of your chat model. Driving a screen means a round trip per look-and-act cycle, so this is the setting where an expensive model adds up fastest.
The iteration cap — default 50, adjustable 10 to 200 — is the brake on that. It is a checkpoint rather than a hard stop: hitting it asks you whether to continue. Lower it while you are learning what these tasks look like. See Letting an agent use your computer.
Teaching an agent runs modules, scenarios and marking, all of which are calls. Teach things you will reuse, not one-offs. See Teaching an agent.
Conversational voice streams replies as they generate and bills your speech provider for both directions.
Meeting transcription does not cost you provider money — it runs on your own machine. The only cost is the one-off model download. See Meeting mode.
Failures cost the same as successes
A task that errors has usually already spent most of what a successful one would. That makes the ERRORS counter on the dashboard a spending figure as well as a health one.
Retry once, then read. The first retry is cheap and often works. A third attempt at an instruction that has failed twice is paying to be told the same thing again — the problem is the instruction, the file path or the key, not luck. See When something does not work.
Watch for an autonomous agent in a loop. An agent that retries on its own can spend steadily while you are not looking. This is one of the practical reasons to promote agents slowly — see Working with agents safely.
A sensible setup
- Connect one provider to start with.
- Set a mid-range model as the default.
- Drop to a small model for the jobs in the top rows of the table, by hand, until you know which they are.
- Keep the computer-use cap low until you have watched a few runs.
- Check your provider console weekly for the first month, then monthly.
If your bill surprises you, the order to check is: computer-use runs first, then teaching, then everything else. The first two are where the volume is.
