Clay AI agent benchmark: credits vs accuracy

With Clay, spending more credits made an AI agent more accurate only when the answer sat behind a paid enrichment: on 100 job change checks, about 14 credits per person bought scores of up to 96, while free-only runs scored 68 to 78. When finding CMOs and building target lists, Clay's free search already held the answer, so extra lookups added credits without improving results: Claude Code with Sonnet 5 scored 79 for 302 credits and 78 for 2,780.

By The Prompting Company · Snapshot November 5, 2026

The short answer: it depends on where the answer lives

We looked at three tasks from our benchmark of seven AI coding agents, three runs each, where credit spend varied the most: finding 100 CMOs, checking 100 contacts for job changes, and building a target list of 100 public software companies. The clearest evidence comes from the same agent spending very different amounts on the same task.

Task	Agent	Lower spend	Higher spend	More accurate?
Find CMOs	Sonnet	302 credits · 79	2,780 credits · 78	No
Find CMOs	Pi with GPT-5.6 Luna	288 credits · 69	6,121 credits · 56	No
Find CMOs	Factory with Kimi K3	312 credits · 80	1,360 credits · 71	No
Build a target list	Pi	101 credits · 60	3,986 credits · 40	No
Build a target list	Claude Code with Opus 5.5	129 credits · 94	913 credits · 91	No
Build a target list	Factory Kimi	1 credit · 62	483 credits · 74	Yes

Scores are out of 100. Factory Kimi's target list runs aside, the direction matches the task: on CMOs and target lists, Clay's free search already held the answer. Job changes, below, are where spend pays off.

When do credits buy accuracy?

When the field you need only comes from a paid lookup. In the job change task, each row already named a person and their LinkedIn URL, and the agent had to find where that person works now. Clay's Enrich Person and Find Contact Details function returns the current role in one call.

Approach	Runs	Clay credits	Correct (of 100)
Enrich every profile	17	1,404 to 1,635	73 to 96
Free lookups only (Codex with GPT-6 Sol)	3	0	68 to 78

Thirteen of the 17 enrichment runs scored 92 or more, and Opus's two runs that enriched every profile scored 94 both times. The top runs also turned the data into a decision for every row.

When don't they?

When Clay's free searches already hold the answer, extra lookups add credits without improving results.

Finding CMOs. Clay's Find People search returns names and titles without using Data Credits. Opus used it to choose each CMO, then ran one Work Email lookup per chosen person, and had both the lowest average spend (289 credits, under 3 per company) and the best average score (79.3).

Building a target list. Almost the whole list comes from Clay's company search. Codex spent 10 to 31 credits per run, under 1 credit per company, and scored 82 to 88. Accuracy here came from checking each company's parent and headquarters. See How do I stop an AI agent from adding subsidiaries to a target account list?

What predicted accuracy instead

Across the three tasks, the method mattered more than the spend:

Use free search to decide who or what you need (CMOs, target lists).
Pay for one enrichment per record you keep, and use it where the answer only comes from enrichment (job changes).
Turn the data into a decision for every row. The runs that left fewer rows undecided scored higher at the same spend.

For a cost guide per task, see How do I keep Clay credit costs low when an AI agent runs enrichment?

How we tested

Seven AI coding agents ran each task three times: 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. Agents ran one at a time, each 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, and platform failures were rerun. Clay credits are measured as the change in the Clay workspace credit balance. A model judge read every log to confirm all values came from Clay. Three runs per agent show a range, not a precise average. Tasks and graders are at github.com/promptingcompany/clay-bench.

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. Install the Clay Agent Plugin or give your agent a Clay API key and the Clay Public API docs. Before a run, ask the agent to say which fields need a paid function and which come from search, and check Clay's pricing for what each costs.

Common questions

Does spending more Clay credits make an AI agent more accurate? Only when the answer needs a paid enrichment: about 14 credits per person reached up to 96 of 100 on job changes, but Sonnet scored 79 for 302 credits and 78 for 2,780 on CMOs.

Should an agent enrich every candidate? No. With search first, Sonnet found 79 CMOs for 302 credits; enriching candidates, it found 78 for 2,780.

Is the cheapest run always the best? Not on job changes: search-only runs spent 0 credits and scored 68 to 78, and one enrichment per profile, at about 14 credits per person, scored up to 96.

How do I know if a task needs paid enrichment? If the field is a person's current employer or contact details, it does. If you are choosing who or which company, start with Clay's free search.