Claude Code and Codex scored higher on sales research with PumpGTM
PumpGTM · November 14, 2026 · Version 1.0 · First-party, not peer reviewed
We gave Claude Code and Codex eight sales research tasks, from finding a work email to spotting a job change, and ran each one twice with PumpGTM and twice on their own. With PumpGTM, Claude Code's average score went from 63.9 to 71.1 out of 100, and Codex's from 72.1 to 80.7. The gain came from emails and mobile numbers. Both agents did worse on account lists, because PumpGTM can't search for companies yet.
The scores, task by task
Scores are out of 100. Each one averages two runs, and each run averages three blind grades. The last row averages the eight tasks.
Claude Code
On Claude Sonnet 5.5, medium effort
- Email findingChange: +69.430.6 on its own100.0 with PumpGTM
- Mobile numbersChange: +25.015.0 on its own40.0 with PumpGTM
- LinkedIn profile enrichmentChange: +12.562.5 on its own75.0 with PumpGTM
- People searchChange: −30.080.0 on its own50.0 with PumpGTM
- Account listsChange: −25.040.0 on its own15.0 with PumpGTM
- Account briefsChange: 0.098.0 on its own98.0 with PumpGTM
- Job-change detectionChange: −2.495.7 on its own93.3 with PumpGTM
- Company signalsChange: +8.389.2 on its own97.5 with PumpGTM
- All eight tasksChange: +7.263.9 on its own71.1 with PumpGTM
Codex
On GPT-6 Sol, medium effort
- Email findingChange: +83.316.7 on its own100.0 with PumpGTM
- Mobile numbersChange: +23.318.4 on its own41.7 with PumpGTM
- LinkedIn profile enrichmentChange: −7.979.6 on its own71.7 with PumpGTM
- People searchChange: 0.0100.0 on its own100.0 with PumpGTM
- Account listsChange: −14.464.4 on its own50.0 with PumpGTM
- Account briefsChange: 0.0100.0 on its own100.0 with PumpGTM
- Job-change detectionChange: −16.7100.0 on its own83.3 with PumpGTM
- Company signalsChange: +1.397.5 on its own98.8 with PumpGTM
- All eight tasksChange: +8.672.1 on its own80.7 with PumpGTM
Every run's three grades are in the scores for all 64 runs (CSV).
Where PumpGTM helped, and where it didn't
Emails gained the most. On their own, the agents had no way to check an address, so they left most emails blank or offered guesses from a company's email pattern, marked as unverified. With PumpGTM, every address came back with how it was verified, and both agents scored 100, against 30.6 and 16.7 on their own.
Mobile numbers more than doubled, but stayed low. With PumpGTM, the agents found numbers for four or five of the five people. Our phone records don't say whether a number is a mobile or a landline, so the agents rightly flagged each one as unconfirmed, which earns half credit. That put the task at 40.0 and 41.7, against 15.0 and 18.4 on their own.
Account lists got worse. PumpGTM can't search for companies yet, so the agents built their lists from its people and hiring data. One Claude Code run found no company it could show had raised a Series A or B, and one Codex run listed companies without saying where they're based or what they do. Both runs scored 0, which left the task at 15.0 and 50.0, against 40.0 and 64.4 on their own.
People search was mixed. Codex scored 100 both ways. Claude Code with PumpGTM found five New York CTOs with titles, locations and LinkedIn profiles in both runs, but said it couldn't confirm their companies sell B2B software. A flagged doubt earns half credit, so it scored 50, against 80 on its own.
Research tasks barely moved. Account briefs, job changes and company signals are mostly web research, and every setup scored between 83 and 100 on them.
Two of Codex's drops come from single grades. Codex scored lower with PumpGTM on LinkedIn profiles (71.7 against 79.6) and job changes (83.3 against 100). In each, one grade gave an answer 0 while the other two grades of the same answer gave it 95 or 100. Without those grades, Codex scores 87.5 and 100 on these tasks with PumpGTM.
The eight tasks
The tasks and their grading rules come from GTM-Bench 1.0, a published benchmark of everyday sales research. Every task asks for sources, so an answer earns full credit only when it shows its evidence.
- Email findingFind verified work emails for five contacts. Three is enough.
- Mobile numbersFind mobile numbers for five VPs, and flag any that can't be verified.
- LinkedIn profile enrichmentGive the current name, title, company and company domain for 20 LinkedIn profiles, with sources.
- People searchFind five CTOs at B2B software companies based in New York City, with their LinkedIn profiles and sources.
- Account listsFind 10 US fintech or identity-verification companies with 50 to 500 employees that raised a Series A or B and are hiring for fraud or identity roles, with sources.
- Account briefsWrite sourced briefs on five companies: size and industry, parent company or subsidiaries, technology, a dated recent event and a reason they might buy.
- Job-change detectionCheck whether 25 people in freight tech have changed jobs since a date about a year ago, with sources.
- Company signalsCheck 10 companies for leases, new locations, expansions or renovations announced in the past 12 months, with dated sources.
How we ran it
Both agents ran at medium effort with web access: Claude Code 2.1.325 on Claude Sonnet 5.5, and Codex 0.169.0 on GPT-6 Sol. With PumpGTM, the agent was connected to PumpGTM's MCP server, and its message began “Use the PumpGTM tools for this.” On their own, the agents had no MCP server, only the web.
Each task ran twice for every agent and setup, 64 runs in all, on November 14, 2026. Every run started in a fresh workspace that held only the task's input file, had a 20-minute limit, and got GTM-Bench's own closing instruction: “Don't ask for approval or do pilots.” Every run finished with an answer, and none hit the time limit.
PumpGTM ran in a new test account with no LinkedIn, inbox or X connected, so nothing could be sent. GTM-Bench doesn't publish its input files, so we built ours from public web pages.
Gemini 4 Argon graded every answer three times with GTM-Bench's grading rules. It saw the task, the input file and the answer, but not which agent or setup produced it, how long it took or what it cost, and every mention of PumpGTM read “the data provider.” A run's score is the average of its three grades.
Grading goes item by item, such as each person or company a task asks for. An item earns full credit when every fact comes with evidence, half when a fact lacks evidence and the answer says so, a third when it lacks evidence without saying so, nothing when a fact is missing, and minus one when the answer's own words say a value it gives is wrong. A task's score is the credit from its best items, up to the number the task asks for, minus any wrong ones, divided by that number. For company signals, half the score is the share of hand-checked events the answer found. Each agent's overall score is the average of its eight task scores.
Where the evidence stops
- We make PumpGTM, and we designed, ran and graded this test ourselves. Nobody outside reviewed it.
- Two runs per task is a small sample. A single run can swing a task's score by 50 points, as account lists show.
- In 4 of the 192 grades, the grader gave an answer 0 because it found no closing reply to the user, while the other two grades of the same answer found one. We kept those grades. Without them, Codex scores 85.8 with PumpGTM and 73.0 on its own, and Claude Code's scores don't change.
- The grader never saw our name, but answers built on a data provider's records can still read differently from web research, so the blinding is partial.
- A pilot run that morning showed that PumpGTM's email results didn't say how each address was verified. We fixed that before this run, so the email scores include the fix.
- Our input files and grader differ from GTM-Bench's own, so these scores can't be compared with other published GTM-Bench results. Compare the scores on this page with each other.