GPT-5.4 is here, and on paper it’s OpenAI’s most direct challenge to Claude yet—“designed for professional work,” better benchmarks, and a big push around agents, computer use, and cleaner knowledge-work outputs. In this video, I walk through the press release, what actually changed versus GPT-5.2, and the pricing that matters if you’re picking a default model for automations. The headline: the context window stays the same (1M), and the big surprise is how expensive output tokens get—especially if you accidentally pick GPT-5.4 Pro. Then I stop reading marketing and run real tests. I put GPT-5.4 head-to-head with Claude Sonnet 4.6 on niche knowledge questions (high-end baseball cards and old presidential card sets) and a spreadsheet cleanup task with messy emails, inconsistent dates, and broken fields. In my tests, Sonnet 4.6 was consistently more accurate on the niche research and did a cleaner job on the spreadsheet—plus it produced an issue log and an easy-to-download cleaned file. GPT-5.4, honestly, hallucinated hard on the presidential set question and didn’t fully deliver on the spreadsheet cleanup. If you’re choosing a model for workflows, this is why you test with your edge cases, not benchmarks.