OpenAI drops GPT-4.5: Bigger, smoother, but not a reasoning beast

OpenAI is framing its newest model as a different kind of step forward. GPT-4.5, released today as a research preview, isn’t being pitched as a raw reasoning monster in the vein of o1 or o3. Instead, the company calls it their largest and most knowledgeable general-purpose model yet, built by scaling up the pre-training that already powered GPT-4o. The goal wasn’t to ace math olympiads. It was to make the thing feel less like a machine and more like a smart colleague who actually gets what you’re asking for.

Early testers seem to agree. The broader knowledge base and what OpenAI describes as improved emotional intelligence translate into a noticeably more natural conversational rhythm. It’s better at writing, better at programming, and crucially, it hallucinates less. That last point alone addresses a persistent thorn in the side of enterprise adoption. The training recipe is familiar territory—a mix of new supervision techniques layered on top of the standard SFT and RLHF methods that have become table stakes in the industry.

On the safety front, the news is refreshingly boring, and that’s exactly how OpenAI wants it. The company’s Preparedness Framework scorecard shows GPT-4.5 docking at Medium risk for CBRN (chemical, biological, radiological, nuclear) threats and persuasion, while cybersecurity and model autonomy sit comfortably at Low. No Critical ratings anywhere on the board. OpenAI explicitly states they didn’t find any significant safety risk increase compared to existing models, which is the kind of non-event safety teams celebrate.

But there’s a reason this is a research preview and not a full-throated launch. OpenAI admits they’re still mapping the model’s edges—what it’s unexpectedly good at, where it quietly stumbles. They’re curious to see how people break it, bend it, and build with it. That candor is rare and useful. The real verdict on GPT-4.5 won’t come from a system card or a benchmark table. It’ll come from developers discovering, weeks from now, that their code just works on the first try more often, or that the model caught a nuance in their prompt they didn’t even realize they’d included.