Bring-your-own-key has one genuine downside: the cost is now yours to watch. Subscription tools hide spend behind a flat fee. With your own key, a careless afternoon of agentic work shows up on a provider invoice, often days later, aggregated into a number you can't attribute to anything.
So Vexi tracks it inline.
Ask at any point
› /usage
12,340 tokens (8.1k in / 4.2k out) · ~$0.0187Input and output are split because they're priced differently: output typically costs several times more per token. When a session gets expensive, the cause is almost always output volume, and the split shows you that directly.
You also get a summary when the session exits, so cost isn't something you have to remember to check.
Why input tokens grow faster than you expect
The counter-intuitive part of agentic cost is that input dominates. Every turn resends context: the project map, the conversation, the file contents the agent read, the tool outputs it received. A twenty-turn session can send the same file twenty times.
This is why the Context Compression Engine is a cost feature as much as a quality one. Compressing old context into a short set of decisions keeps the resent payload small. Without it, long sessions get quadratically more expensive as they go.
Free models read as zero, honestly
Known free-tier models report $0 rather than a notional price. This is deliberate, because showing "what this would have cost on GPT-4o" would be a fictional number, and a cost display that reports fiction is worse than no display.
If you're on Groq's free tier, your cost is zero, and that's what it says.
It's an estimate, and here's the gap
The dollar figure is calculated locally from token counts and a table of per-model prices. That's an estimate, not your invoice, and it can drift from reality in a few specific ways worth knowing:
- Prices change. The table ships with the release. If a provider cuts prices tomorrow, your local estimate is stale until you update.
- Prompt caching isn't reflected. Providers that discount repeated context will bill you less than the estimate shows.
- Unknown models can't be priced. A custom endpoint or a model missing from the table yields token counts without a dollar figure: a blank rather than a guess.
Treat it as a live gauge, not an accounting record: excellent for noticing that a session is running hot while you can still do something about it, not a substitute for your provider's billing page at the end of the month.
What to do with the number
The practical value is calibration. Most developers badly misjudge AI cost in both directions: some avoid a useful tool over pennies, others burn real money on a loop they forgot was running. A week of watching /usage fixes both, and after that you'll mostly stop looking.