Clay AI agent benchmark: batching and credit cost

With Clay, batch size did not change the credit cost: AI agents that sent the same 100 contacts to Clay's Work Email routine as ten batches of 10, two batches of 50, or a test of 5 and then 95 each paid exactly 337.5 Clay credits. Across 21 runs by seven AI coding agents, the bill followed how many items were sent to Clay, not how they were grouped.

By The Prompting Company · Snapshot November 5, 2026 · Three runs per agent

What did the same contacts cost in different batches?

We asked seven AI coding agents to find work emails for 100 contacts using Clay's Public API and a saved Work Email routine. Each agent chose its own batching, which gave us a clean comparison: the same contacts, the same routine, different batch sizes.

How the agent batched its 100 contacts	Agent and run	Clay credits
One test contact, then ten batches of 10	Pi, all three runs; Grok, runs 1 and 2	337.5
One test contact, then two batches of 50	OpenCode Kimi, run 1	337.5
One test contact, a pilot of 5, then 95	Factory Kimi, run 2	337.5

Clay credits are measured as the change in the Clay workspace credit balance. All of these runs found the same 93 verified emails.

What does change the cost?

The number of items sent. In runs that used only the Work Email routine, each extra item sent to Clay added about 3 credits, whatever the batch size:

Items sent to Work Email	Batching	Clay credits
100	One test, then 25, 25, 25 and 24	334.5
101	Ten batches of 10, two of 50, or 5 then 95	337.5
102	Two tests, then five batches of 20	340.5
103	Tests of 1 and 2, then one batch of 100	343.5

The rest of the spread came from optional steps, not batching. Codex with GPT-6 Sol, which split its work into many small batches in all three runs, spent 335.5 credits when it went straight to Work Email. Clay's pricing page states that an enrichment with no result is not charged, so the lowest-cost runs sent each contact once and left empty rows blank.

Does batch size change the speed?

In these runs, larger batches finished sooner. Clay runs each batch asynchronously and the agent polls for results. A batch is done when its last item returns (agents logged batches at 9 of 10, 19 of 20 or 24 of 25 just before that), so fewer batches meant fewer polling rounds.

Minutes are measured from run start to grading.

Batching	Run	Minutes
A test, a pilot of 5, then 95	Factory Kimi, run 2	4.8
A test, then two batches of 50	OpenCode Kimi, run 1	5.1
A test, then ten batches of 10	Pi, runs 1 to 3	8.9 to 12.1
A test, then ten batches of 10	Grok, runs 1 and 2	10.1 and 12.4

This is an observation from a small number of runs, not a controlled test of speed, and agents also differ in how often they poll. Cost, by contrast, matched to the half credit.

What is the best batching pattern for an AI agent in Clay?

The fastest clean runs followed the same shape:

Send one test item first. Clay's API told every agent exactly which fields the routine needs (Full Name, Company Domain, Company Name) in its response to an early call, so agents self-corrected in one call. A one-contact test costs a few credits.
Send the rest in one or two large batches. Batch size doesn't change the bill, and fewer batches meant fewer polling rounds.
Save progress as results arrive, so a slow batch doesn't hold up the file.
Leave empty rows blank. Clay doesn't charge when an enrichment returns nothing, so one pass per contact is all the routine needs.

The full cost picture, including what each agent paid per email found, is in "How many Clay credits does it cost an AI agent to find a work email?"

How we tested

Seven agents ran the task three times each: Claude Code with Opus 5.5 and Sonnet 5, Codex with GPT-6 Sol, Pi with GPT-5.6 Luna, Grok CLI with Grok 4.6, OpenCode with Kimi K3 and Factory with Kimi K3. Each run was alone in a fresh sandbox with the same Clay key, the same 19 saved Clay routines and a 60-minute limit; a timeout would count as 0, and runs stopped by a platform failure were rerun. All 21 runs finished. Because agents ran one at a time, each run's credits are the exact change in the workspace balance, and we matched them to the batch sizes and item counts Clay recorded for each routine run.

This is one task with one routine. Routines that call different data providers will cost different amounts per item, but the same rule applies: count the items, not the batches.

Try it yourself

The task and grader are at github.com/promptingcompany/clay-bench. Install the Clay Agent Plugin, or give your agent a Clay API key and the Clay Public API docs, and tell it to send one test item, then the rest of the list in one or two batches, and not to retry empty rows. Compare your workspace credit balance before and after to check the cost. The Clay Agent Plugin ships Clay's skills to your agent up front, including the habits that keep credit use down; setup is one prompt (about 15 minutes), and there is no additional cost to use it.

Common questions

Does batching Clay API calls change the credit cost? No. The same 100 contacts cost exactly 337.5 credits in batches of 10, in batches of 50, and as 5 then 95.

What does change the Clay bill? The number of items sent, at about 3.4 credits per contact for Work Email in our runs. Each extra Work Email item added about 3 credits, and running extra lookups, such as a company lookup on every row, adds credits without improving results.

Is a bigger batch faster? In our runs, one or two large batches finished in about 5 minutes and ten batches of 10 took 8.9 to 12.4, measured from run start to grading, because each batch is done when its last item returns.

Should an agent send a test call first? Yes. It costs a few credits, and Clay's response tells the agent exactly which inputs the routine needs before the full run.