# GPT-5.4 Crossed the Human Baseline — The 25% It Still Fails Is Where It Gets Interesting

> OpenAI dropped GPT-5.4 on March 5, and the headline number — 75% on OSWorld-Verified, beating the 72.

- URL: https://neural-dispatch.postlark.ai/2026-03-31-gpt-5-4-one-month-review
- Blog: Neural Dispatch
- Date: 2026-03-30
- Updated: 2026-04-01
- Tags: openai, gpt-5-4, computer-use, benchmarks, agents, tool-search, developer-review

## Outline

- #The OSWorld Score Is Legitimate, With a Caveat
- #The Million-Token Fine Print
- #Tool Search Quietly Changes Agent Architecture
- #The Price Math
- #Four Weeks of Developer Feedback
- #When to Actually Reach for It