GPT-5 Is Here: What Changed, What Didn't, and Why It Matters Less Than You Think

The Launch Nobody Expected

OpenAI released GPT-5 on a Tuesday afternoon with a blog post and a shrug. No keynote, no countdown timer, no Sam Altman tweet storm. After years of increasingly theatrical AI launches, the quiet release felt almost subversive.

But the model itself is anything but quiet. GPT-5 represents a genuine step change in reasoning, code generation, and multi-modal understanding. The question is whether that matters as much as it would have two years ago.

What Actually Improved

The benchmarks are impressive. GPT-5 scores 94.2% on MMLU (up from 86.4% for GPT-4), handles 200K context windows natively, and produces code that passes unit tests on first attempt 78% of the time (compared to 52% for GPT-4 Turbo).

More practically: it follows complex instructions more reliably, hallucinates less frequently, and can maintain coherent reasoning across document-length outputs. It's the first model that can reliably summarize a 300-page legal document without losing the thread.

What Didn't Change

GPT-5 still can't reliably do math beyond what it's seen in training data. It still sometimes confuses correlation with causation in its reasoning. It still generates plausible-sounding nonsense when pushed beyond its knowledge boundaries — it's just better at admitting it doesn't know.

And critically, it still requires human oversight for any high-stakes application. "Better" is not "trustworthy." The error rate dropped from ~15% to ~6% on complex reasoning tasks, but that 6% can still be catastrophically wrong.

Why It Matters Less Than You Think

Here's the uncomfortable truth: for most people, GPT-4 was already "good enough." The jump from GPT-3.5 to GPT-4 changed what was possible. The jump from GPT-4 to GPT-5 mostly changes what's convenient. The ceiling of AI capability went up, but the floor of user expectations went up faster.