Clay AI agent benchmark: the cheapest agent to run

With Clay, the cheapest workflows in our benchmark cost about 3.6 Clay credits per correct work email, about 3.6 per CMO found and under 1 per qualifying company on a target list. Across all 15 of its runs, Codex with GPT-6 Sol spent the fewest Clay credits, 3,937, about 2.6 per record, while Pi with GPT-5.6 Luna had the lowest model cost, $0.25. No agent was cheapest on both bills, and the agent's workflow moved the Clay bill far more than the choice of agent.

By The Prompting Company · Snapshot November 5, 2026

What does it cost to run an agent with Clay?

There are two bills: Clay credits for the data, and the model cost of running the agent. Clay prices in Actions (platform usage) and Data Credits (vendor data and AI), as set out on Clay's pricing page. In this benchmark, Clay credits are measured as the change in the Clay workspace credit balance. Two Clay rules kept agent costs down: there is no additional cost to use the Clay Agent Plugin, and if an enrichment returns no result, you're not charged Data Credits or Actions.

Model cost is as reported by the test platform. It was not reported for the three runs that reached the 60-minute limit (one by Claude Code with Opus 5.5 and two by OpenCode with Kimi K3, called OpenCode Kimi below), so those agents' totals below leave those runs out.

Which agent was cheapest overall?

Totals across all 15 runs (5 tasks, 3 runs each), ordered by Clay credits.

*Model cost not reported for timed-out runs: one for Opus, two for OpenCode Kimi.

Codex's total includes three job change runs that used Clay's search and spent 0 credits. Pi's model cost was $0.01 to $0.03 per run.

Which agent was cheapest per correct result?

A cheap run only helps if it is right. This table shows the lowest average Clay credits per correct result on each task. Where that agent's average score was below the task's pass line (marked †), the next row shows the most efficient agent whose average cleared it.

†Average score below the task's pass line: 90 on domain and HQ for company data, 80 for job changes, 70 for the target list.

On work emails, Grok CLI with Grok 4.6 came close to Pi at 339 credits for 93 emails, at $0.65 of model cost per run instead of $0.01. On company enrichment, Codex's extra 85 credits over Grok bought about 15 more correct domains. On job changes, OpenCode Kimi decided nearly every row for about 15 credits per correct answer. On the target list, Codex listed 86 qualifying companies on average for 20 credits, and Grok 65 for 11.

What kept the Clay bill low?

Every task could be done for a few credits per record or less. The lowest-cost runs chose each record with Clay's free search, then ran one paid routine for it.

In Clay, sourcing a list with Find People uses 0 Data Credits, so the agent pays only for the records it keeps. More in "How do I keep Clay credit costs low when an AI agent runs enrichment?".

What about the model bill?

Model cost ranged from $0.01 per run for Pi to $4.44 for the most expensive CMO run by Claude Code with Sonnet 5. For the other six agents, the median run cost $1.04, so Pi's total for all 15 runs, $0.25, was less than 86 of their 87 reported runs. Long runs cost more: the first enrichment run by Factory with Kimi K3 took 48.9 minutes and cost $3.70.

How we tested

Seven agents ran each of five tasks three times: Claude Code (Opus 5.5 and Sonnet 5), Codex (GPT-6 Sol), Pi (GPT-5.6 Luna), Grok CLI (Grok 4.6), OpenCode (Kimi K3) and Factory (Kimi K3). Each run had a fresh sandbox, the same Clay key and the same 19 saved Clay routines, and a 60-minute limit; a timeout counts as 0. Agents ran one at a time, so each run's Clay credits are the exact change in the workspace balance. Platform failures were rerun and are not counted. Tasks, prompts and graders are at github.com/promptingcompany/clay-bench.

What this doesn't tell you. Clay credits here mix Actions and Data Credits as one balance, so we do not convert them to dollars; your rate depends on your Clay plan. Model prices change often.

Try it yourself

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. Give your agent the Clay Agent Plugin or a Clay API key and the Clay Public API docs. Tell it to use Clay's free people or company search to choose each record, then run one paid routine per chosen record. Check the workspace credit balance before and after a test batch of 10 to see your cost per result.

Common questions

What is the cheapest AI agent to run with Clay? Codex with GPT-6 Sol used the fewest Clay credits, 3,937 across 15 runs or about 2.6 per record, and Pi with GPT-5.6 Luna had the lowest model cost, $0.25 across 15 runs.

How many Clay credits does an agent need per result? About 3.6 per correct work email or CMO in the cheapest runs, and under 1 per qualifying company on a target list.

Does it cost extra to use Clay from an AI agent? No. There is no additional cost to use the Clay Agent Plugin, and an enrichment that returns nothing is not charged.