RealRun / Results

Which AI can schedule a meeting without double-booking you?

Tested on 62 realistic meeting requests (42 main + 20 fresh, held-back) across 8 AI models, 2+ passes each: 1,116 graded answers.

Meeting scheduling test, October 2026. 42 hand-written scheduling requests plus 20 fresh holdout requests written blind, 8 AI models, the same frozen prompt.

Short answer: Claude Opus, GPT-6.1, Gemini 3.8 Flash and Grok 4.7 got every item right on both sets. Claude Sonnet and Claude Haiku were close behind. GPT-5.4 mini slipped on hard cases, and Grok 4.20 got only about 4 in 10 right, so we don't recommend it.

Scorecard

Validity = share of items answered correctly: a time that fits the request, the calendar and working hours, or correctly saying no time works or that a request is a scam. Average of 2-3 passes per model. On the 20-item holdout, one item is 5 points.

The top four hit the ceiling. Opus, GPT-6.1, Gemini Flash and Grok 4.7 scored 100% on both sets, so this test can't rank them against each other. A harder test would be needed.

Safety

• No scam failures. Every model declined every scam request (look-alike senders pushing log-in or payment links).
• No over-declines. No model refused a legitimate Zoom, Meet or Teams invite as a scam.

⚠ Not recommended for scheduling: Grok 4.20. It often proposed the start of the offered window even when that time was busy, said the slot was free, and made time-zone and daylight-saving errors.
⚠ Caution: GPT-5.4 mini on hard cases. On the fresh holdout it sometimes booked over a busy event or offered a time when none worked. Check every time it proposes.

Most misses by the other models were saying "no time works" when one did, the safe direction. See the playbook for item-by-item misses and per-pass scores.

What the prompt does

• Uses the latest message: if people change the date, length or options, the newest request wins.
• Checks the whole calendar: recurring meetings, skipped dates, all-day and tentative events, and working hours in your time zone.
• Handles time zones and daylight saving, including ambiguous ones, and states the zone in the reply.
• Picks the earliest time that works, or says politely that none does.
• Safety and privacy: declines scam requests without repeating their links, never claims it already booked anything, and never reveals your other meetings.
• Returns a draft reply plus the time in a format you can check.

An excerpt, as proof it's a real, specific prompt:

You are scheduling meetings for [your name] ([your email]). Their time zone is [zone]. ...
Below are their calendar (an ICS-style export) and an email thread asking for a meeting. ...
Output only this JSON object, with no other text: ...

What you get in the $29 playbook (instant PDF download):

• The exact, word-for-word tested prompt, with a fill-in template.
• Full results with per-pass scores, cost and reply-quality checks for all 8 models.
• Every miss, item by item, so you know where each model slips.
• A setup guide (marked untested) and safety rules.

Buy the scheduling playbook — $29 (founding price)

How we tested

• Each item: one email thread, the user's calendar as an ICS export, and the user's time zone, working hours and current time.
• Main set of 42 hand-written items: time zones and daylight saving, relative dates, mind changes, double-booking, working hours, scams, legitimate video invites and ambiguous zones. Holdout of 20 items written blind by a separate worker and used once.
• One item per request, the same frozen prompt, temperature 0 where allowed. 2 passes per set, with a 3rd when passes differed by more than 4 points.
• Objective scoring: a time counts only if it has the right length, sits inside an offered time, is within working hours and doesn't overlap the calendar. Re-scored independently from the raw answers, with hand spot-checks.
• All dummy data. No real inbox or calendar was used.

Caveats

• Synthetic threads and calendars; yours will be messier.
• We tested plain AI models, not scheduling apps or assistants that book for you.
• 62 items in total. One item moves the holdout score by 5 points.

Also see: Which AI can you trust to sort your inbox?

Get the free weekly report