Why is self-scrape per-task cost different from raw cost?

Fixed costs (proxies, servers, engineer time) are divided by usable successes, not attempts. If your success rate is 85%, you need 1,176 attempts to get 1,000 usable tasks, so the effective cost per usable task is higher than the naive division.

How should I amortize fine-tuning training cost?

A common heuristic is 12 months of useful life. Take total training cost (compute + data labelling) and divide by 12 to get a monthly amortized number, then add per-call inference on top.

Does Sella include payment fees?

Yes. The Sella per-call price is gross to you as a buyer. Platform fees and on-chain gas are deducted from the publisher side. See the x402 Pricing Estimator for the seller view.

Tools

Agent cost comparison: buy live data, self-scrape, or fine-tune?

Three patterns for getting data into your agent. This calculator scores all three on monthly total cost, cost per usable task, and accuracy, so you can pick the one that fits your volume, accuracy bar, and ops bandwidth.

Tasks per month

Buy on Sella

Price per call (USDC)

Self-scrape

Proxy network / mo

Servers / mo

Engineer hrs / mo

Engineer $ / hr

Scrape success %

Fine-tune

Training amortized / mo

Inference $ / call

Model accuracy %

Approach	Monthly total	Cost / usable task	Accuracy
Buy on SellaBest 100% pay-per-success. No infra. No model risk.	$1.00K	$0.0100	100%
Self-scrape Fixed-cost. 85,000 usable / mo at current success rate.	$1.24K	$0.0146	85%
Fine-tune Training amortized over 12 mo + per-call inference.	$1.70K	$0.0170	75%

Recommendation

Buy on Sella: lowest cost per usable task, zero infra to maintain.

Browse marketplace →

When each pattern wins

Buy on Sella wins at low-to-mid volumes, when you need fresh data, or when the accuracy bar is high. You pay only for successful calls. Self-scrape wins at large stable volumes against unprotected sources, but it carries success-rate risk, legal exposure, and ongoing maintenance. Fine-tune wins when the knowledge is durable (won't go stale in weeks) and the accuracy ceiling of your base model is sufficient.

Once you've picked: browse the Sella marketplace for the dataset or API you need, or run the Agent Spend Calculator to project monthly burn at your chosen price point.

FAQ

Why is self-scrape per-task cost different from raw cost?: Fixed costs (proxies, servers, engineer time) are divided by usable successes, not attempts. If your success rate is 85%, you need 1,176 attempts to get 1,000 usable tasks, so the effective cost per usable task is higher than the naive division.
How should I amortize fine-tuning training cost?: A common heuristic is 12 months of useful life. Take total training cost (compute + data labelling) and divide by 12 to get a monthly amortized number, then add per-call inference on top.
Does Sella include payment fees?: Yes. The Sella per-call price is gross to you as a buyer. Platform fees and on-chain gas are deducted from the publisher side. See the x402 Pricing Estimator for the seller view.

Sella, the economy for AI agents

Discover, preview, and pay for machine-payable datasets, evals, and APIs on the Sella marketplace. Or publish your own.