OpenAI’s GPT-5.2 Gamble Pays Off Big

OpenAI released GPT-5.2 today and it totally outshines the competition. Working behind the scenes, a crew of engineers at OpenAI’s San Francisco HQ spent months adjusting and tweaking this beast of a model. Its results speak for themselves, as it aced every single test they’ve run so far.

The new model comes in three variants, each designed to tackle a specific job. The Instant version is great for quick lookups and casual conversations, having no problems extracting info from big datasets. But when you hit the tough stuff, Thinking kicks in like a personal project manager, breaking down math proofs or knocking out a timeline for your next big project one step at a time. If accuracy is paramount, the Pro version is where it’s at for those really tricky tasks.

When it comes to benchmark tests, the numbers don’t lie. On the GDPval exam, which pits the AI against some of the world’s top human specialists in 44 different fields, the Thinking model is coming out on top 70.9% of the time. How long did those same jobs take human workers to finish? Well, the AI zipped through them in a fraction of the time and for a fraction of the cost. It’s not a one off either, as other challenges are showing the same results, with GPT-5.2 nailing 80% of its software engineering tasks (up from 76.3 in the last version) and then just breezing through some math questions that even its predecessors had struggled with, scoring a perfect 100%.

Vision tasks got a significant upgrade this time around. The model is much better at reading charts and images to say the least, spotting trends in data graphics, as well as navigating all sorts of app interfaces. It’s also cool under pressure, no matter how long a document is. GPT-5.2 can plough through hundreds of thousands of words without breaking a sweat, summarizing reports or contracts with near 100% accuracy. Of course, how it interfaces other tools is important, whether you’re uploading a spreadsheet or chaining a series of actions together in a process. Even large projects, such as handling a consumer complaint from beginning to end, run smoothly.

This time around, safety got a bit more attention than usual. OpenAI has managed to knock 30% off false claims, so you can be pretty sure your answers are on the money. Conversations concerning delicate topics, such as mental health or personal concerns, receive more thoughtful responses, and the results suggest that the model is performing better in these situations than ever before. A new age checker is in the works to keep an eye on younger users, and next year we’ll be rolling out an adults-only setting that’ll let people open up more without worrying about it being seen by others.

The new versions will be available to all paid ChatGPT subscribers, regardless of what level they’re on (Plus, Pro, Go, Business, and Enterprise). They’re calling these models GPT-5.2, and if you’re one of those subs you’ll be getting the Thinking build for now, with Instant and Pro versions following soon.