GPT-6 Sol or GPT-5.6 Luna: which OpenAI model works better with Clay?
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 AI agent benchmark: GPT-6 Sol vs GPT-5.6 Luna
With Clay, GPT-6 Sol in Codex and GPT-5.6 Luna in Pi both found 93 of 100 verified work emails in every run, and each fits a different job. Codex had the highest average score of seven agents, 83.2 of 100, on the fewest Clay credits, and built qualifying target lists in under 5 minutes for 10 to 31 credits. Pi was the fastest agent overall and the lowest-cost model to run, $0.25 for all 15 runs, and it checked 100 contacts for job changes in 4.7 minutes.
By The Prompting Company · Snapshot November 5, 2026
The question
Two of the seven setups in our Clay benchmark ran OpenAI models: Codex with GPT-6 Sol and Pi with GPT-5.6 Luna. Both the model and the harness differ, so this compares the two setups as you would run them, on the same five GTM data tasks with Clay, three runs each. Clay credits are measured as the change in the Clay workspace credit balance, and minutes from run start to grading.
Task by task
Primary score out of 100, averaged over three runs; Clay credits and minutes are the median run.
| Task | Codex | Pi | Codex credits | Pi credits | Codex minutes | Pi minutes |
|---|---|---|---|---|---|---|
| Work emails | 93.0 | 93.0 | 426 | 337.5 | 10.1 | 9.3 |
| Company data | 90.5 | 66.0 | 341 | 933 | 9.3 | 4.8 |
| CMO | 73.3 | 63.0 | 405 | 304 | 9.1 | 10.2 |
| Job changes | 73.0 | 80.7 | 0 | 1,404 | 15.1 | 4.9 |
| Target list | 86.0 | 33.3 | 18 | 167 | 4.3 | 5.3 |
| All five | 83.2 | 67.2 | 1,190 | 3,146 | 47.9 | 34.5 |
| Codex · GPT-6 Sol | Pi · GPT-5.6 Luna | |
|---|---|---|
| Runs passed, of 15 | 8 | 5 |
| Model cost, 15 runs | $14.38 | $0.25 |
| Model cost per run | $0.38 to $2.11 | $0.01 to $0.03 |
Where GPT-6 Sol in Codex does best
Target lists. Codex built its lists from Clay's company search in 4.2 to 4.8 minutes for 10 to 31 credits and listed 88, 88 and 82 qualifying companies, checking that each was listed in its own name and headquartered in the US.
Company matching. Codex averaged 98.3 on domain, HQ country and headcount, the best of any agent, by checking the candidates in Clay's company search before enriching.
The fewest Clay credits. Codex spent 1,313 credits for one average run of all five tasks, about 2.6 credits per record, the lowest of seven agents.
Work emails and CMOs. Codex found 93 of 100 work emails in every run, and in its third run went straight to Work Email for 335.5 credits, about 3.4 per contact. Its first CMO run found 81 of 100.
Where GPT-5.6 Luna in Pi does best
Work emails for about a cent. Pi ran Work Email once per contact and found 93 of 100 in every run for exactly 337.5 credits, the lowest average on the task, at $0.01 of model cost per run. With Clay's Work Email waterfall, you only pay credits for the provider that finds a match.
Speed. Pi's median runs of all five tasks added up to 34.5 minutes, the lowest of seven agents, with the fastest median on company data, 4.8 minutes, and on job changes, 4.9.
Job changes. Pi ran Enrich Person and Find Contact Details on all 100 LinkedIn profiles, about 14 credits per person, and scored 84 and 85 in two runs. The 85 took 4.7 minutes.
The lowest model bill. Pi's $0.25 for all 15 runs was the lowest model cost in the benchmark; 86 of the other six agents' 87 reported runs each cost more than that on their own.
Running extra lookups added credits without improving results, so naming one paid routine per record in the prompt keeps either setup's Clay bill low.
Which one fits which job?
| Job | Good fit | Why |
|---|---|---|
| Work emails at scale | Pi | 93 of 100 for 337.5 credits and $0.01 of model cost |
| Target lists | Codex | 82 to 88 qualifying for 10 to 31 credits |
| Company matching and enrichment | Codex | 98.3 on domain, HQ and headcount |
| Quick job change checks | Pi | 100 profiles in under 5 minutes, 85 of 100 |
| Lowest model bill | Pi | $0.25 for 15 runs |
| Lowest Clay bill across all five tasks | Codex | 1,313 credits per average run |
Both OpenAI setups did strong work with Clay; the right pick depends on the task. For cost across all seven agents, see "What is the cheapest AI agent to run with Clay?", and for Pi on its own, "What can Pi do with Clay?".
How we tested
Seven agents ran each of five tasks three times, one agent 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, and platform failures were rerun. Model cost is as reported by the test platform. Tasks, prompts 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 in Codex or Pi, or give either a Clay API key and the Clay Public API docs. Name the Clay method in the prompt: Work Email directly for emails, company search with an ownership and HQ check for target lists, and one enrichment per profile for job changes. Agent work uses the same credits as in-product work (Clay pricing).
Common questions
GPT-6 Sol or GPT-5.6 Luna: which OpenAI model works better with Clay? Each fits different jobs: GPT-6 Sol in Codex had the highest average score, 83.2 of 100, on the fewest Clay credits, and GPT-5.6 Luna in Pi was the fastest and had the lowest model cost, $0.25 for 15 runs.
Which is better for work emails with Clay? Both found 93 of 100 in every run; Pi did it for 337.5 credits and about 1 cent of model cost per run.
Which is better for target lists? Codex: 82 to 88 qualifying companies per list for 10 to 31 credits, in under 5 minutes.