RealRun / Guides
Short answer: GPT-6.1 Sol was at or near the top in all three of our tests: 95% on inbox triage, 100% on scheduling and 95.8 on receipts. Claude Opus was close behind on email and scheduling. Older or cheaper Grok models (4.20, 4.3) had the most serious problems.
| Model | Inbox triage (%) | Scheduling (%) | Receipts (score) |
|---|---|---|---|
| GPT-6.1 Sol | 95.0 | 100.0 | 95.8 |
| GPT-5.4 mini | 95.0 | 90.0 | — |
| GPT-6 Luna | — | — | 91.4 |
| Claude Opus | 94.4 | 100.0 | 89.7 |
| Claude Sonnet | 91.7 | 100.0 | — |
| Claude Haiku | 85.6 | 96.7 | 92.8 |
| Gemini 3.8 Flash | 91.1 | 100.0 | — |
| Gemini 4 Argon | — | — | 93.6 |
| Gemini 3.5 Flash-Lite | — | — | 85.4 |
| Grok 4.7 | 90.0 | 100.0 | 83.1 |
| Grok 4.20 | 86.7 | 41.7 | — |
| Grok 4.3 | — | — | 69.2 |
— = not tested on that task. Each test used 8 models. Headline metrics differ by task; see each results page.
Get the tested prompts. $29 playbooks: Inbox triage · Meeting scheduling · Receipts
Scores reflect our tests on the dates shown, not a guarantee. Claude models ran on a Claude subscription, so they have no metered API cost. cost_tier is relative within a task (low/medium/high) from our run spend; 'subscription' = not metered. Tested October 2026 by RealRun (Base Ledger IP LLC).