Skip to main content
Effort sets how many angles the agent tries, how many sources it reads, and how long it may run. Higher is not better by default: it buys sources, rows and time, and past a point it buys nothing the task needs. Fix the task and the schema first; raise effort last.

​
What a tier actually buys

Typical numbers from a recent month of external runs: Read across a row, not down a column:
  • Research: medium returns a complete free-text brief. high reads five times the sources and fills every cell when a schema is attached. x-high doubles time again for a few more sources and rows.
  • Enrichment: high fills nine cells in ten when the batch fits the budget. Size the batch to the fields: five rows with six researched fields finished in 4.5 minutes; ten rows with the same fields ran 21 minutes and researched only five. x-high fills a little more and halves partial output, but failed more often on model timeouts.
  • Dataset building: x-high returns about a quarter more rows than high at four times the price. Scope beats effort: one run over a whole manufacturer catalog at x-high returned 78 items; scoped runs at the same tier with the traversal written into the skill returned more than 200 each.
Plan for the tail. Nine in ten high and x-high runs finish in under 20 to 30 minutes; a synchronous caller needs a timeout at least that long.

​
Find the cheapest tier that holds

The pattern that works in customer pilots:
1

Pick 100 representative inputs

Include the hard cases and a few you already know the answers for.
2

Run them at the highest tier you would pay for

This is your quality ceiling for this task and schema.
3

Run the same 100 one tier lower

Compare filled cells, rows returned and spot-checked accuracy, not the confidence grade.
4

Keep the cheapest tier whose results you would ship

On a 100-item product-matching pilot, high matched x-high at a quarter of the price. On a 10,000-company financial enrichment, high filled 96.7% of cells with 70% of rows complete.

​
Cost arithmetic

A job costs rows ÷ rows per run × price per run, plus the runs that fail or need a top-up. Two limits shape the batch column. Enrichment runs accept at most 20 input rows and 200 input cells (rows × fields), and the practical batch is smaller: what one run finishes at the chosen effort. Rows beyond the budget come back with only their input values. Accounts run 30 concurrent agents and 200 runs per day by default; contact Nimble ahead of a larger job.

​
Preview before committing

Run one batch, or one region, before the whole job. A first run at the recommended tier on a sample of 10 tells you whether the task and schema hold. A low run does not: it reads a handful of pages and returns thin output for anything beyond a fact check, so it previews the agent’s understanding of the ask, not the quality of the result.
max (a custom budget per run) is not yet available. Requests for max return 422 today.

​
Next Steps

Efforts

Tiers, times and prices

Running at Scale

Batches, concurrency, retries

Evaluating Results

What to compare between tiers