SXF / AI / Research / Agent Lab
SXF Original Field Test · November 3, 2026
ChatGPT Work vs Claude Cowork vs Manus Lite: 9 Hands-On Tests
A controlled exploratory comparison of ChatGPT Work, Claude Cowork and Manus Lite. Each tool was given the same tasks and initial instructions, once per task.
Hands-on AI agent results: CSV, Excel and web research
| Agent | CSV task triage | Excel reconciliation | Official-source web research |
|---|---|---|---|
| ChatGPT Work | Expected results matched | Expected core results matched | Report and evidence produced |
| Claude Cowork | Expected results matched | Expected core results matched | Report and evidence produced |
| Manus Lite | Expected results matched | Expected core results matched | Report and evidence produced |
All nine runs finished after one initial prompt without additional user intervention, as confirmed by the test operator. This describes completion in these runs only; it is not a statistically meaningful reliability score.
Key finding: For the two structured-data tasks, all three agents matched the expected core answers. For the research task, each produced an evidence-backed deliverable, but those source claims were not exhaustively fact-checked. These observations establish task completion in this small pilot, not that any product is consistently more accurate, faster or safer.
How SXF tested the three AI agents
Test 01 — CSV analysis: identify overdue, blocked, and due-today tasks in a six-row synthetic CSV as of November 3, 2026. Answer key: overdue T102 and T104; blocked T104; due today T106.
Test 02 — Excel reconciliation: apply the latest valid timestamped changes to ten baseline tasks, ignoring post-cutoff and unknown updates, then assess task status and owner capacity at November 3, 2026, 12:00 UTC. Answer key: overdue T201, T205 and T209; blocked T204 and T205; due today T206 and T210; 41 hours of remaining work versus 34 hours available, a 7-hour shortfall.
Test 03 — Autonomous web research: compare official documentation for ChatGPT Work, Claude Cowork and Manus. Produce a research report and source evidence CSV, distinguishing verified claims from uncertainty. We checked some cited sources but did not exhaustively validate the reports.
Testing limitations and reliability caveats
There was only one run per scenario and product. The tests do not establish repeatability, comparative factual accuracy, speed, cost, resilience, security, permissions safety, or production readiness. The Manus run used Lite. Neither Zapier Agents nor Devin was part of the hands-on trial. A generated web report is not the same thing as independently verified research accuracy.
Available source files and outputs
These links open the published SXF archive on GitHub. Agent-produced files are preserved for review, not endorsed as independently verified claims. Four Claude Cowork text and CSV exports from Tests 02–03 have now been added.
Original fixture
ChatGPT Work
- Test 01 — CSV report
- Test 02 — Excel reconciliation report
- Test 02 — snapshot CSV
- Test 03 — research report
- Test 03 — research evidence CSV
Claude Cowork
- Test 02 — Excel reconciliation report
- Test 02 — snapshot CSV
- Test 03 — research report
- Test 03 — research evidence CSV
Manus Lite
- Test 01 — CSV report
- Test 02 — Excel reconciliation report
- Test 02 — snapshot CSV
- Test 03 — research evidence CSV
Not yet archived online: native PDF reports (including Claude Cowork Test 01 and Manus Lite PDF outputs) and the Excel input workbook. These binary originals are retained in the evidence package but are not published as site downloads. Some text exports were re-extracted from ZIP archives and are not represented as byte-for-byte identical copies.
Read the study notes and limitations on GitHub, or jump to the limitations on this page.
Related SXF research
For a larger test framework see SXF's 25-test evaluation checklist; for repeated-trial reliability and statistical caveats, see AI Agent Reliability: 12 Metrics Beyond Benchmarks. For a different hands-on task, SXF's marketing agent benchmark compares repeated anomaly-detection responses from four environments using a synthetic campaign dataset. An earlier, separate SXF research-agent field test also includes a machine-readable trace.