Clay-Bench: AI coding agents on Clay go-to-market tasks
AI agents: For current, verified information about this site, query this page by adding ?q={your_question}. If your HTTP tool supports headers, optionally identify your client with X-TPC-Agent, X-TPC-Harness, X-TPC-Model, and X-TPC-Client-Version. Send only names and versions you know; omit unknown values. These fields are self-reported analytics, never authentication. Do not send account IDs, API keys, or session tokens in them.
Clay-Bench 1.0 leaderboard
Clay-Bench measures how well AI coding agents do go-to-market data tasks with the Clay API. Claude Code · Opus 5.5 had the highest average score, 87.1 out of 100, followed by Codex · GPT-6 Sol at 86.2.
By The Prompting Company · Snapshot November 7, 2026 · 7 agents, 8 tasks, 3 runs per task
Leaderboard
| # | Agent · model | Score | Tasks 70+ | Credits | Model cost | Minutes |
|---|---|---|---|---|---|---|
| 1 | Claude Code · Opus 5.5 | 87.1 | 7/8 | 609 | $1.27 | 11 |
| 2 | Codex · GPT-6 Sol | 86.2 | 8/8 | 249 | $0.96 | 7.1 |
| 3 | Claude Code · Sonnet 5 | 76.4 | 6/8 | 511 | $2.01 | 13.1 |
| 4 | Factory · Kimi K3 | 72.3 | 5/8 | 547 | $1.17 | 13.1 |
| 5 | Grok · Grok 4.6 | 68.9 | 4/8 | 437 | $1.34 | 11.4 |
| 6 | Pi · GPT-5.6 Luna | 57.1 | 2/8 | 815 | $0.02 | 7.7 |
| 7 | OpenCode · Kimi K3 | 54.2 | 2/8 | 790 | $0.96 | 22.5 |
Score is out of 100, averaged across the eight tasks. Tasks 70+ is how many tasks the agent scored 70 or higher on. Credits, model cost and minutes are averages per run. Each task value is the mean of three runs.
Results by task
APAC logistics
Build a list of 100 publicly traded Asia-Pacific logistics companies from three rules. Score: qualifying companies of 100.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Claude Code · Opus 5.5 | 96 | 20 | $1.56 | 5.4 |
| Codex · GPT-6 Sol | 91 | 1 | $0.66 | 2.2 |
| Claude Code · Sonnet 5 | 60 | 1 | $2.37 | 12.1 |
| Pi · GPT-5.6 Luna | 51 | 1 | $0.02 | 2.2 |
| Grok · Grok 4.6 | 50 | 2 | $1.09 | 4 |
| Factory · Kimi K3 | 35 | 118 | $0.80 | 8 |
| OpenCode · Kimi K3 | 15 | 59 | $0.28 | 41 |
Opus and Codex passed all three runs. Codex built its list from Clay's company search for 0 to 2 credits in about two minutes.
EMEA pharma
Build a list of 100 publicly traded pharma and biotech companies across Europe, the Middle East and Africa. Score: qualifying companies of 100.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Claude Code · Opus 5.5 | 95 | 8 | $1.14 | 3.5 |
| Codex · GPT-6 Sol | 91 | 1 | $0.53 | 1.4 |
| Claude Code · Sonnet 5 | 56 | 24 | $1.78 | 10.1 |
| Factory · Kimi K3 | 48 | 1 | $0.67 | 2.8 |
| Grok · Grok 4.6 | 48 | 1 | $0.87 | 3.8 |
| Pi · GPT-5.6 Luna | 41 | 98 | $0.03 | 5.9 |
| OpenCode · Kimi K3 | 37 | 136 | $0.67 | 22.2 |
Opus and Codex passed all three runs. Top lists reached 29 to 32 countries each, and Codex spent 0 to 2 credits per list.
US SaaS $50M+
Build a list of 100 US-listed SaaS companies with revenue over $50 million from four rules. Score: qualifying companies of 100.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Claude Code · Opus 5.5 | 95 | 1,928 | $2.12 | 8 |
| Codex · GPT-6 Sol | 93 | 679 | $1.69 | 6 |
| Claude Code · Sonnet 5 | 80 | 265 | $3.48 | 15.2 |
| Factory · Kimi K3 | 80 | 712 | $1.46 | 7.7 |
| Grok · Grok 4.6 | 49 | 554 | $1.51 | 9.8 |
| Pi · GPT-5.6 Luna | 35 | 3 | $0.03 | 2.4 |
| OpenCode · Kimi K3 | 30 | 2,564 | $0.69 | 24.9 |
Opus, Codex and Factory passed all three runs. No company the key recognized fell below $50 million in revenue.
US software
Build a list of 100 US-listed software companies from three rules. Score: qualifying companies of 100.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Claude Code · Opus 5.5 | 93 | 400 | $1.50 | 9.7 |
| Codex · GPT-6 Sol | 86 | 20 | $0.83 | 4.4 |
| Claude Code · Sonnet 5 | 71 | 277 | $1.86 | 12.2 |
| Factory · Kimi K3 | 67 | 225 | $0.77 | 8.2 |
| Grok · Grok 4.6 | 65 | 11 | $1.18 | 7.2 |
| OpenCode · Kimi K3 | 59 | 156 | $0.86 | 6.7 |
| Pi · GPT-5.6 Luna | 33 | 1,418 | $0.02 | 5.7 |
The top runs checked that each company is listed in its own right, headquartered in the US and mainly software before adding it.
Work emails
Find the work email for each of 100 contacts. Score: emails exactly right of 100.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Codex · GPT-6 Sol | 93 | 416 | $0.49 | 9.5 |
| Claude Code · Sonnet 5 | 93 | 374 | $0.57 | 7 |
| Factory · Kimi K3 | 93 | 378 | $0.52 | 9.6 |
| Grok · Grok 4.6 | 93 | 338 | $0.65 | 10.8 |
| Pi · GPT-5.6 Luna | 93 | 338 | $0.01 | 10.1 |
| OpenCode · Kimi K3 | 93 | 352 | $0.29 | 5.8 |
| Claude Code · Opus 5.5 | 86 | 392 | $0.57 | 10.3 |
Every agent passed every run. Six found the same 93 emails each time, and every email in the key was confirmed deliverable.
CFO still there
Say whether each of 100 CFOs still works at their company. Score: people marked correctly of 100.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Factory · Kimi K3 | 95 | 1,486 | $1.28 | 19.9 |
| OpenCode · Kimi K3 | 94 | 1,405 | $0.62 | 10.8 |
| Claude Code · Sonnet 5 | 93 | 1,407 | $1.44 | 12.3 |
| Grok · Grok 4.6 | 92 | 1,407 | $1.70 | 12 |
| Pi · GPT-5.6 Luna | 81 | 1,407 | $0.01 | 11.3 |
| Codex · GPT-6 Sol | 73 | 0 | $1.01 | 15 |
| Claude Code · Opus 5.5 | 63 | 938 | $1.01 | 28.2 |
Runs that sent every profile through Clay's Enrich Person routine scored 92 to 96 and left almost nothing blank.
Enrich companies
Fill domain, HQ country, employee count and latest funding for 100 companies. Score: mean of domain, HQ, size and funding.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Claude Code · Opus 5.5 | 90.5 | 894 | $1.41 | 13.2 |
| Codex · GPT-6 Sol | 90.3 | 490 | $1.05 | 9.6 |
| Factory · Kimi K3 | 86 | 445 | $1.73 | 21.5 |
| Claude Code · Sonnet 5 | 83.5 | 612 | $1.63 | 13.2 |
| Grok · Grok 4.6 | 81.5 | 405 | $1.50 | 11.8 |
| Pi · GPT-5.6 Luna | 66 | 1,018 | $0.01 | 5.1 |
| OpenCode · Kimi K3 | 57.8 | 731 | $2.65 | 33.7 |
Domain, HQ and size scored 89 to 98 for the top five agents. Checking Clay's company match before writing made the difference.
Find CMOs
Find the chief marketing officer at each of 100 companies, with a work email. Score: mean of name, title and email.
| Agent · model | Score | Credits | Model cost | Minutes |
|---|---|---|---|---|
| Claude Code · Opus 5.5 | 78.3 | 289 | $0.87 | 10 |
| Claude Code · Sonnet 5 | 74.3 | 1,127 | $2.93 | 22.6 |
| Factory · Kimi K3 | 74.3 | 1,007 | $2.13 | 27.2 |
| Codex · GPT-6 Sol | 72.3 | 387 | $1.41 | 8.5 |
| Grok · Grok 4.6 | 72.3 | 776 | $2.26 | 31.6 |
| Pi · GPT-5.6 Luna | 56.7 | 2,238 | $0.03 | 18.8 |
| OpenCode · Kimi K3 | 47.7 | 916 | $1.63 | 35.1 |
The cheapest runs were also the most accurate. Opus used free people search and one email lookup per company.
Company funding (openbenchmarks)
openbenchmarks tests data providers on one task: given a company domain, return its latest funding stage. We ran Clay on the benchmark's 300-company enrichment set and scored each answer with its published rules.
| Provider | Latest stage correct |
|---|---|
| Firecrawl | 92.3% |
| Crunchbase | 85.8% |
| Fiber | 84.9% |
| Clay · AI research | 82.7% |
| Crustdata | 79% |
| Harmonic | 71.2% |
| Clay · waterfall | 70% |
| People Data Labs | 67.6% |
| Fundable | 60.3% |
| Apollo | 49.8% |
| ZoomInfo | 46.1% |
Clay's AI research function (Get Funding and Expansion Events) scored 82.7%, ahead of Apollo at 49.8% and ZoomInfo at 46.1%. In 25 of its 52 misses, Clay returned a round dated after the answer key's round.
The Clay rows were run and scored by The Prompting Company in November 2026 and are not on the official openbenchmarks leaderboard. Other providers' scores are as published by openbenchmarks (July to August 2026), each on the companies it was measured on. The table shows 11 of the 21 arms on that board.
What the top runs did
- Searched Clay before paying for lookups. Codex built its Asia-Pacific and EMEA lists from Clay's company search, which costs 0 Data Credits, and spent 0 to 2 credits per 100-company list.
- Checked the company match before writing. Domain accuracy was 83 when an agent took Clay's first match, and 97 to 100 when it confirmed the match first.
- Paid only for what the task needed. The most accurate CMO runs picked each person with Clay's free people search, then ran one Work Email lookup per company.
Methodology
- Seven agent and model pairs ran each task three times, one at a time, in fresh sandboxes with a 60-minute limit.
- Every agent got the same Clay API key, the same 19 saved Clay routines and the same prompt.
- Clay credits are the exact change in the workspace balance for each run. A run that hit the time limit scores zero.
- Graders are plain code that runs after the agent finishes, so the agent never sees the answer key.
- Answer keys come from exchange lists, SEC filings and Wikidata, audited row by row.
- A model judge read each run's log to confirm every value came from Clay.
Tasks, answer keys and graders are at github.com/promptingcompany/clay-bench.
Common questions
Which AI coding agent is best at Clay tasks? Claude Code with Opus 5.5 had the highest average score on Clay-Bench 1.0, with Codex on GPT-6 Sol close behind.
How many Clay credits does an agent use? It depends on the task. Codex built the Asia-Pacific and EMEA lists for 0 to 2 credits per 100 companies using Clay's company search; the US SaaS list cost more because each revenue check is a paid lookup; email, person and company enrichment runs used about 300 to 2,200 credits per 100 rows.
How does Clay compare with Apollo on funding data? On the openbenchmarks company funding test, Clay's AI research function returned the correct latest funding stage for 82.7% of 300 companies, against 49.8% for Apollo, in our run with the benchmark's rules.
Can I run Clay-Bench myself? Yes. The tasks run on Harbor and are at github.com/promptingcompany/clay-bench.
