OpenAI Released GPT-5
By Jeremy ·
GPT-5 is live as of today, available through ChatGPT and the API. OpenAI describes it as "unified intelligence and action in one model," which is the kind of sentence that gets written when marketing is in the room. The practical change is the end of the two-track model setup they've been running since the o-series launched.
One Model Instead of Two
For the past year, OpenAI's lineup required a choice between GPT-4o for fast, general-purpose responses and the o-series models for extended reasoning tasks. That split created friction for developers and confused users about when to use which.
GPT-5 handles both within a single call. The model decides internally whether a query warrants a quick response or a longer reasoning pass. You don't specify a mode or pick a model variant. That means simpler API integrations for anyone building on top of it, and a cleaner experience in ChatGPT where the reasoning/non-reasoning distinction was never intuitive to most people.
The Benchmark Picture
The performance numbers are the strongest OpenAI has published at this point.
On AIME 2025 (competitive mathematics), GPT-5 scores 94.6% without any tool use. On SWE-bench Verified (real-world software engineering tasks), 74.9%. On Aider Polyglot (coding across multiple languages), 88%. On MMMU (multimodal reasoning across images and text), 84.2%.
On factual accuracy: when web search is enabled, GPT-5 responses are about 45% less likely to contain a factual error compared to GPT-4o. That's a meaningful improvement for anyone using the model for research or content work where accuracy matters.
Full details are on OpenAI's GPT-5 page.
What This Changes for Operators
The practical shift for anyone building on the OpenAI API is simpler model selection. The previous setup meant deciding which tasks needed o3, which needed GPT-4o, and wiring up two different integrations. GPT-5 consolidates that.
Multimodal input is standard rather than a separate capability. Images, documents, and text go into the same model without any preprocessing routing.
For publishers and editorial teams using AI tools, the hallucination reduction is probably the most useful change. The jump from GPT-4o's accuracy floor to GPT-5's is significant enough that it changes what workflows are reliable versus which ones still require a human checkpoint on every output.
The o-series models don't disappear immediately, but their distinct positioning gets harder to justify now that a single model handles the reasoning-intensive tasks they were built for. What that means for OpenAI's model roadmap going forward is worth watching.