Independent lab scores OpenAI Dots 8.8/10 on real business tasks — behind Claude Cowork and ChatGPT Work

#What happened

An independent test lab (aiagentslibrary.com, no sponsorships, failures included) gave five AI agents — Claude Cowork, ChatGPT Work, OpenAI Dots, Gemini Spark and Meta Muse — a run sheet of 10 comparable real business tasks each, scored on the same five criteria between August and October 2026. It is the first head-to-head measurement of Dots against its agent rivals on real work.

The measured averages: Claude Cowork 9.3/10, ChatGPT Work 9.15/10, OpenAI Dots 8.8/10, Gemini Spark 8.7/10. Meta Muse, tested September 27 on the free plan, earned a 6.5/10 overall impression (8.0 average across its eight scored tasks; four planned tasks went unscored after it hit password or Full Disk Access requests the reviewer would not approve).

Dots was tested September 30 to October 1, 2026 — the day after launch — on a ChatGPT Pro account with Google Workspace, GitHub, ChatGPT Sites, Figma, Adobe, Stripe and Slack connected. Its strongest results were four tasks at 9.8/10: accurate platform research, publishing a page through ChatGPT Sites from one prompt, a content plan from real business context, and a recurring scheduled job. Its weakest was post-meeting follow-up at 7/10, and its clearest failure was a spreadsheet that reported passing 27 formula checks and then broke on import to Google Sheets — the reviewer notes "a self-reported pass is not verification."

The reviewer also reported a separate 9.2/10 overall experience rating for Dots, reflecting the always-on experience (scheduling, phone notifications, Slack, voice calls) that the five criteria do not score.

#Why it matters

This is one of the first independent, task-by-task measurements of Dots as a working agent rather than a feature demo — and it lands third of five. For buyers comparing always-on agents, the spread is tight at the top (9.3 vs 9.15 vs 8.8), and the lab's verdict on Dots' spreadsheet failure is a useful caution: judge agents on imported, usable output, not on their self-reported checks.

Sources: https://www.aiagentslibrary.com/ai-agent-test-lab/ — Verified October 23, 2026.