Which AI agent finishes Clay tasks fastest?
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: the fastest agent
With Clay, the fastest AI agent in our benchmark, Pi with GPT-5.6 Luna, finished a typical run of each of five data tasks in 34.5 minutes combined, measured as the sum of its median times. Codex with GPT-6 Sol was second at 47.9 minutes and built a target list of 88 qualifying companies in about 4 minutes. Most runs of 100 records finished in under 15 minutes; speed came from letting one Clay routine or search do the work for each record.
By The Prompting Company · Snapshot November 5, 2026
Which agent was fastest on each task?
Median minutes across three runs per task, measured from run start to grading. This covers more than the agent's own working time, so these times are longer than the agent-only times in the work email post. A run that reached the 60-minute limit counts at the time it stopped.
| Agent | Work emails | Enrich companies | CMO | Job changes | Target list | Sum of medians |
|---|---|---|---|---|---|---|
| Pi · GPT-5.6 Luna | 9.3 | 4.8 | 10.2 | 4.9 | 5.3 | 34.5 |
| Codex · GPT-6 Sol | 10.1 | 9.3 | 9.1 | 15.1 | 4.3 | 47.9 |
| Claude Code · Opus 5.5 | 7.8 | 12.9 | 9.2 | 14.2* | 8.5 | 52.6 |
| Claude Code · Sonnet 5 | 7.1 | 15.3 | 22.6 | 12.1 | 11.9 | 69.0 |
| Factory · Kimi K3 | 8.7 | 7.9 | 23.2 | 21.0 | 9.0 | 69.8 |
| Grok CLI · Grok 4.6 | 10.1 | 12.1 | 29.1 | 11.7 | 7.3 | 70.3 |
| OpenCode · Kimi K3 | 5.1 | 31.0* | 27.6* | 10.3 | 6.4 | 80.4 |
*One of the three runs reached the 60-minute limit and counts as 0.
Pi had the lowest median on two tasks, Codex on two, and OpenCode with Kimi K3 on work emails, at 5.1 minutes from run start to grading for 100 contacts.
Which agents were fast and accurate?
Codex is the clearest example of fast and accurate together. It built its target lists in 4.2 to 4.8 minutes and listed 88, 88 and 82 qualifying companies. It scored 98 to 99 on company domains, HQ and headcount in 7.2 to 12.2 minutes. Claude Code with Opus 5.5 found the most CMOs, 79.3 on average, at a median of 9.2 minutes.
| Agent | Task | Median time | Average score |
|---|---|---|---|
| Codex | Target list | 4.3 min | 86 |
| Codex | Enrich companies (domain) | 9.3 min | 98.3 |
| Opus | CMO | 9.2 min | 79.3 |
| Claude Code with Sonnet 5 | Job changes | 12.1 min | 93.3 |
| OpenCode Kimi | Work emails | 5.1 min | 93 |
What made runs fast?
The fast runs had two things in common.
- One routine per record. On work emails, the agents ran Clay's Work Email routine for each contact they searched, and the fastest median was 5.1 minutes. On job changes, the agents that ran Enrich Person and Find Contact Details once per LinkedIn profile finished in 4.7 to 24.4 minutes.
- Search instead of enrichment for list building. Codex built its target lists almost entirely from Clay's company search, spending 10 to 31 credits, in under 5 minutes each time.
How long did the CMO task take?
The CMO task took the longest for most agents. Opus, which searched first and ran Work Email only for the chosen person, finished each CMO run in 8.0 to 12.8 minutes. One routine per record kept job change runs short too: Opus's runs that used one enrichment per profile finished in 9.8 and 14.2 minutes. For the cost side of the same choices, see "What is the cheapest AI agent to run with Clay?".
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 task had 100 records. 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, and Clay credits are measured as the change in the Clay workspace credit 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. Each run started from scratch, with only the prompt, Clay's docs and the saved routines, so time covers working out the approach as well as running it. Three runs show a range, not a precise median.
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. Name the routine to use for each record and tell the agent to run it once per record. For a list of 100, allow at least 30 minutes; 78 of the 102 completed runs in our benchmark finished in under 15.
Common questions
Which AI agent finishes Clay tasks fastest? Pi with GPT-5.6 Luna: a sum of 34.5 median minutes across five Clay tasks of 100 records each.
Which agent was fast and accurate? Codex with GPT-6 Sol. It was second fastest at 47.9 minutes and built an 88-company target list in about 4 minutes.
How long does an AI agent take to find 100 work emails with Clay? The fastest median was 5.1 minutes from run start to grading, by OpenCode with Kimi K3 (3m 41s of agent time, as the work email post reports), and every run finished within 18.8 minutes.
How long should I allow for a list of 100? At least 30 minutes; 78 of the 102 completed runs in our benchmark finished in under 15.