Clay AI agent benchmark: which AI coding agent gets the most out of Clay

With Clay as the constant, Claude Code with Opus 5.5 got the most out of it: it passed 14 of its 15 runs across five sales data tasks, more than any other agent. On average score the top four agents finished within one point of each other, 82.3 to 83.2 out of 100. The Clay method each agent chose on each task explained most of the gaps.

By The Prompting Company · Snapshot November 5, 2026

The question

We ran seven AI coding agents through five GTM data tasks with Clay, three runs each: find work emails, enrich company data, find each company's CMO, check contacts for job changes, and build a target account list. Clay and its saved routines stayed the same for every run, so the differences come from the agent and its approach.

The overall leaderboard

Each task has one primary score out of 100: emails correct (task 1), the average of domain, HQ, headcount and funding (task 2), CMO name correct (task 3), people correctly marked as moved or not (task 4) and qualifying companies (task 5). Each cell is the average of three runs. Overall is the plain average of the five task scores. A run passes when it clears every pass line for its task.

Agent	Emails	Company data	CMO	Job changes	Target list	Overall	Runs passed
Codex · GPT-6 Sol	93.0	90.5	73.3	73.0	86.0	83.2	8 of 15
Claude Code · Sonnet 5	93.0	83.5	75.0	93.3	70.7	83.1	10 of 15
Factory · Kimi K3	93.0	85.9	74.7	94.7	66.7	83.0	11 of 15
Claude Code · Opus 5.5	86.0	90.7	79.3	62.7	92.7	82.3	14 of 15
Grok CLI · Grok 4.6	93.0	81.5	73.0	92.3	65.0	81.0	8 of 15
OpenCode · Kimi K3	93.0	58.0	48.3	93.7	58.7	70.3	9 of 15
Pi · GPT-5.6 Luna	93.0	66.0	63.0	80.7	33.3	67.2	5 of 15

Three of 105 runs reached the 60-minute limit and count as 0 in these averages.

Opus led three of the five tasks, and its two finished job change runs scored 94. Codex had the top average and spent the fewest Clay credits, measured as the change in the Clay workspace credit balance: 1,313 for one average run of all five tasks, about 2.6 credits per record across 500 records. Pi had the lowest model cost, $0.25 across all 15 runs.

Which Clay method worked best on each task?

The same method gave close to the same result whichever agent used it.

Task	Method in the top runs	Result
Work emails	Work Email routine once per contact, no company lookup first	93 of 100 for 338 credits (Pi) and 339 (Grok)
Company data	Check the candidates in Clay's company search before enriching; take funding from company news	Codex 98 on domain, HQ and headcount; Opus 75 on funding
CMO	Free people search to choose the person, then Work Email for that one person	Opus 74 to 82 for 285 to 293 credits
Job changes	Enrich Person and Find Contact Details once per LinkedIn profile, then decide every row	92 to 96 in every run with 3 or fewer rows left undecided
Target list	Clay's company search, checking ownership and headquarters before adding a company	Opus 91 to 94; Codex 82 to 88 for 10 to 31 credits

What the top runs had in common

One enrichment per profile on a full list. On job changes, 17 runs enriched every profile, at about 14 credits per person; 16 of them passed and 13 scored 92 or more.

A short, direct workflow. The top runs named one Clay routine or search per record and ran it, which kept most runs short: 78 of 102 completed runs finished in under 15 minutes.

A check on each company before adding it. On the target list, Opus and Codex confirmed each company was listed in its own name and headquartered in the US, and listed 82 to 94 qualifying companies per run.

For the head-to-head detail, see "Claude Opus or Sonnet: which is better at using Clay?" and "Codex or Claude Code: which works better with Clay?".

How we tested

Seven agents (Claude Code with Opus 5.5 and with Sonnet 5, Codex with GPT-6 Sol, Pi with GPT-5.6 Luna, Grok CLI with Grok 4.6, OpenCode with Kimi K3, Factory with Kimi K3) ran every task three times, one at a time, in a fresh sandbox with the same Clay key, the same 19 saved Clay routines and a 60-minute limit. A timeout counts as 0; platform failures were rerun.

Pass lines: 70 for emails; 90 for domain and HQ and 70 for headcount and funding; 70 for CMO name and title and 60 for email; 80 for job changes; 70 qualifying companies plus a full list of 100 for the target list. Counting only runs that kept every value from Clay, Opus passed 13 of 15, still the most. Credits per agent are the sum of its five per-task averages.

Try it yourself

All five tasks are published in Harbor format at github.com/promptingcompany/clay-bench. 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. To run your own agent against Clay, install the Clay Agent Plugin or give the agent a Clay API key and the Clay Public API docs, and tell it which method to use from the table above. Agent work uses the same credits as in-product work (Clay pricing).

Common questions

Which AI coding agent gets the most out of Clay? Claude Code with Opus 5.5 passed 14 of 15 runs, the most of seven agents; Codex had the highest average score, 83.2 of 100.

Does the agent matter more than the method? Mostly no. Every agent that used Work Email directly found the same 93 of 100 emails, and every job-change run that enriched each profile and decided nearly every row scored 92 to 96.

Which agent used the fewest Clay credits? Codex, with 1,313 credits for one average run of all five tasks, about 2.6 credits per record.

Which task gave the most consistent results? Work emails: six of seven agents found the same 93 of 100 in every run with Clay's Work Email routine. See "Where should an AI agent get funding rounds in Clay?" for the company data method that led on funding.