OpenAI drops GPT-5.2 with expert-level knowledge work skills

OpenAI just dropped GPT-5.2, rolling out Instant, Thinking, and Pro models aimed squarely at knowledge workers who live in spreadsheets, slide decks, and codebases.

The update is available immediately in the API for developers, while ChatGPT Plus, Pro, Business, and Enterprise users get access starting today. Instead of chasing generic benchmarks, OpenAI is targeting high-value corporate desk work.

Taking over the spreadsheet grind

The headline claim from OpenAI is that GPT-5.2 Thinking beats or matches top industry professionals in 70.9% of comparisons on GDPval, a benchmark covering tasks across 44 different occupations. The company estimates the model completes these tasks at 11 times the speed and less than 1% of the cost of human experts.

For anyone in finance, the improvements are concrete. On an internal benchmark replicating junior investment banking tasks, such as building three-statement corporate models or leveraged buyout projections, GPT-5.2 Thinking scored 68.4%, up from 59.1% in GPT-5.1.

Software engineering also saw big jumps. GPT-5.2 Thinking hit 80% on SWE-bench Verified and 55.6% on SWE-Bench Pro, which spans four languages with stricter anti-contamination measures. Front-end engineers testing early builds reported much cleaner handling of complex layouts and 3D web elements.

Hallucinations dropped by 30% relative to GPT-5.1 Thinking across a sample of real-world queries. Meanwhile, on long-context document analysis, the model achieved near 100% accuracy on 4-needle retrieval tests across 256,000 tokens.

Vision, tools, and graduate-level math

Visual reasoning saw its error rate cut roughly in half when reading charts and software interfaces. In spatial tasks, such as locating components on a blurry motherboard photo, the new model generates reasonably accurate bounding boxes where its predecessor failed.

Tool-calling performance reached 98.7% on Tau2-bench Telecom. In complex customer support scenarios involving flight delays, rebooking, medical assistance, and compensation, the model successfully chained all four steps end-to-end without dropping context.

On the scientific frontier, GPT-5.2 Pro reached 93.2% on GPQA Diamond and became the first model to crack 90% on ARC-AGI-1 Verified. It also solved 40.3% of expert-level problems on FrontierMath.

Why it matters

We think GPT-5.2 represents a decisive shift from chat assistants to actual white-collar workflow execution.

If you are an analyst, engineer, or consultant, this is definitely worth testing right away. The ability to churn out properly formatted financial models and workable front-end prototypes moves AI from an ideation partner to a direct production tool. Enterprise heavy users already report saving ten hours a week, and this release aims to expand that.

However, the trade-offs are obvious. Complex spreadsheet and presentation generations can take several minutes to finish. Crucially, human review is still non-negotiable, as judges noted minor errors still slip through despite the polished layout.

For the AI ecosystem, OpenAI is putting intense pressure on vertical SaaS wrappers. When the foundation model can autonomously coordinate multi-step workflows, manage 256k tokens reliably, and build corporate deliverables from scratch, third-party enterprise tools will have to prove they offer much more than just a slick prompt UI.