Grok vs. Dot vs. Muse vs. Claude Duet: which one crushes the competition?
- ID
- 40357
- Status
- summarized
- Published
- 28 Oct 2026, 8:04 PM
- Fetched
- 28 Oct 2026, 9:51 PM
- Provider
- Lenny's Newsletter
- Category
- product-startup
- Original URL
- https://www.lennysnewsletter.com/p/grok-vs-dot-vs-muse-vs-claude-duet
- Source URL
- https://www.lennysnewsletter.com/feed
Summary
- Score
- 5.5
- Created
- 28 Oct 2026, 9:51 PM
- Tags
- Audience
- ai_agent_usersai_ml_learnersdevelopersvibe_coders
What happened
Claire Vo ran what the episode calls the first 'How I AI' personal-agent benchmark: Grok Bot, Dot, Meta Muse, and Claude Duet each attempted nine real tasks — a Slack reply written in her voice, cutting short-form video from raw Dropbox footage, fixing a bug from Sentry, building a Figma thumbnail, inbox triage, a family newspaper, a fourth-grade physics homework explanation, shoe shopping, and a ChatPRD launch video. The harness was itself agent-operated: Claude ran Dot while Codex ran the other three and scored the results live. The text provided is show notes only — no per-task scores, final totals, or named winner appear in it, so the title's ranking is not substantiated here.
Why it matters
The copyable asset is the harness, not the verdict: nine specific, checkable tasks plus the rule that a different agent operates each contestant is a template you can run against your own two-or-three agent shortlist instead of trusting a headline ranking. The scoring setup is also a caution — Codex graded every agent except the one Claude operated — so before you switch tools off any benchmark, check who ran and who graded the contestants. This excerpt gives you zero numbers to act on, so treat it as a methodology pointer only.
Discussion angle
Have each person name the three tasks they'd actually test a personal agent on, then debate whether a benchmark where Codex grades its rivals and Claude grades Dot can produce a trustworthy winner at all.