GPT-5 arrives with a brain router and a real shot at usefulness

OpenAI pulled the curtain back on GPT-5 this morning, and the headline isn’t just about benchmark scores—it’s about a fundamental architectural bet. The new model isn’t a single monolith. It’s a unified system with a real-time router that decides, moment by moment, whether your query needs a quick, efficient answer or the full weight of a deeper reasoning engine they’re calling GPT-5 thinking. The router trains continuously on real-world signals, like when users manually switch models or thumbs-up a response, which means the system should get smarter about triaging your requests the more you use it.

On paper, the numbers are genuinely state-of-the-art. We’re looking at 94.6% on AIME 2025 for math, 74.9% on SWE-bench Verified for real-world coding, and 84.2% on MMMU for multimodal understanding. The health benchmark score hit 46.2% on HealthBench Hard, which OpenAI published earlier this year with physician-defined criteria. But the more interesting story might be what early testers are saying about the model’s design sensibilities—apparently it’s got a much better eye for typography, spacing, and white space, generating responsive websites and apps from a single prompt that don’t just function but actually look tasteful.

Writing and health are two other pillars OpenAI is pushing hard. For writing, they claim GPT-5 handles structural ambiguity with a literary depth that previous models fumbled, like sustaining unrhymed iambic pentameter without sounding like a robot reading a textbook. On the health side, the model acts more like an “active thought partner,” proactively flagging concerns and adapting its responses based on your context and knowledge level. The company is careful to frame it as a partner for understanding results and weighing options—not a replacement for a doctor—but the specificity of the geography-aware and knowledge-level-adaptive features suggests this is more than just a disclaimer.

If you’re a Plus subscriber, you get more usage. Pro users unlock a version with extended reasoning that hit 88.4% on GPQA without tools. Once you blow through usage limits, a mini version of each model handles the rest. The near-future plan to collapse all this routing intelligence into a single model feels like the real endgame here—a system that doesn’t just answer questions but genuinely knows when to think and when to act.