Models, estimates and the cap
A run spends money on one thing: model calls, billed to your key. This page is about which model, what it is likely to cost, and how the cap keeps it bounded.
The model
Section titled “The model”The Model field on the Start panel suggests the models your key can list, and fills in the first of these that the key offers:
| Provider | Suggested, in order |
|---|---|
| OpenRouter | deepseek/deepseek-v4.1-flash, google/gemini-3.5-flash, anthropic/claude-haiku-4.5 |
| OpenAI | gpt-5-mini, gpt-5, gpt-4.1-mini |
| Anthropic | claude-haiku-4-5, claude-sonnet-4-5 |
gemini-3.5-flash, gemini-2.5-flash |
|
| OpenAI-compatible | none |
If the key offers none of its provider’s suggestions, the first model in its list, alphabetically, is filled in. You can type the exact name of any other model the key can use. The one model you choose does everything in the run: it plays every person, carries out every replay, and judges.
Most of what a run costs is input — the pages the model reads, step after step — so an inexpensive model usually gets further within the same cap than a large one.
Prices
Section titled “Prices”Prices come from OpenRouter’s public price list, refreshed at most once an hour, for every provider — the model names of OpenAI, Anthropic and Google keys are matched to it. The Start panel shows the price under the model, for example $0.30 in / $1.20 out per million tokens. Models on an OpenAI-compatible service never have a price.
The price is fixed when the run starts, so a price list that changes or disappears later never makes a run cheaper to count. Through OpenRouter, a call is counted at the cost OpenRouter reports for it when that is higher than the fixed price.
The estimate
Section titled “The estimate”Estimated run is worked out from fixed token counts, not from your product:
- One step of a person’s session is counted as 10,000 tokens in and 300 out.
- The low end assumes 20 steps per person; the high end assumes every person uses all 60.
- Each defect that gets replayed adds about 15 steps plus one judge call of 6,000 tokens in and 150 out.
So a plan with four people on a model at $0.30 in and $1.20 out per million tokens is estimated at $0.27–$0.81, with 4 people, plus about $0.05 for each defect that gets replayed. Real runs vary with how much text your pages have and how quickly the people get through their goals.
The cap
Section titled “The cap”The Hard cap is where the run stops spending, from $0.10 to $50, $2 unless you change it. Every model call the run makes is counted, at the price fixed when the run started — or, through OpenRouter, at the cost OpenRouter reports for the call, if that is higher.
- The run stops as Stopped at cap when spending reaches the cap. Sessions, replays and judge calls still waiting are cancelled; everything found so far stays in the report.
- A call is only ever allowed as much output as the remaining money pays for, but its cost is counted once it has happened, so the last call can take the total a little past the cap.
- The call Start makes to check the key is not part of the run and is not counted against the cap.
- Judging a defect again after the run is paid from the same cap, so it needs room left under it.
- If a runner is stopped during a person’s session and hands the session back, that person starts again from the beginning. If the runner disappears without handing it back, the session fails after ten minutes with the runner stopped answering. Either way, what was already spent still counts.
Setting up a project — reading the page and proposing the plan — runs on Trawler’s own model and costs you nothing.
When the price is unknown
Section titled “When the price is unknown”If the Start panel has no price to show for the model — none is known, the models could not be listed, or you typed a name that is not in the list — the estimate reads Unknown and the cap field disappears, leaving the default $2.
When the run starts, Trawler looks the price up again. If it finds one, the run counts dollars against the $2 cap as usual. If it does not, the run also stops after 3 million tokens — input and output together — and the run page shows Tokens so far against that cap. Such a run can count dollars only where OpenRouter reports a call’s cost, so watch your provider’s billing.