OpenAI reaches benchmarks highs in programming and professional tasks with new GPT-5.2

The new model is primarily oriented towards the business and professional environment, positioning itself as the most advanced currently available according to various industry benchmarks. OpenAI highlights that GPT-5.2 sets new records in areas such as professional work, software engineering, and extensive context analysis.

The model includes three variants: Instant, Thinking, and Pro. Instant is oriented towards daily use and quick queries, Thinking is designed for complex tasks requiring greater depth of analysis, and Pro offers the highest level of quality for difficult questions where accuracy is prioritized over speed.

In the GDPval benchmark, which evaluates specialized knowledge tasks across 44 occupations, GPT-5.2 Thinking matches or exceeds expert professionals in 70.9% of cases. Tasks include creating presentations and spreadsheets that the model completes at eleven times the speed and less than 1% of the cost compared to specialized human work.

One of the most significant improvements is in programming. GPT-5.2 Thinking reaches 55.6% on SWE-Bench Pro, a benchmark that evaluates the resolution of real software engineering problems in four languages. In financial modeling tasks with spreadsheets, accuracy increases from 59.1% to 68.4%. Test users have highlighted notable improvements in developing complex interfaces with three-dimensional elements.

The model expands its capacity to work with extensive documents, reaching almost 100% accuracy in analyzing information distributed across up to 256,000 tokens, equivalent to several hundred pages. This feature is especially useful for analyzing contracts, technical reports, or projects with multiple files.

In visual processing, the model approximately halves the error rate in interpreting scientific graphs. Accuracy in analyzing professional screenshots increases from 64.2% to 86.3%, facilitating the analysis of dashboards and technical diagrams.

OpenAI reports a 30% reduction in hallucinations compared to the previous version. In expert-level mathematics, GPT-5.2 Thinking solves 40.3% of problems in FrontierMath, compared to 31% for GPT-5.1. The model also improves in coordinating multiple tools, reaching 98.7% accuracy in multi-step customer service tasks.

GPT-5.2 is available starting today in ChatGPT for paid plan users and in the API for all developers. OpenAI has set an API price higher than GPT-5.1 per token, though it remains below other reference models in the market. The company indicates that despite the per-token increase, the final cost to achieve a given quality level is lower due to the new model's greater efficiency.

Key points

• GPT-5.2 sets new records in multiple industry benchmarks
• GPT-5.2 Thinking matches or exceeds expert professionals in 70.9% of specialized work tasks across 44 different occupations
• The model reaches 55.6% on SWE-Bench Pro, setting a new record in solving real software engineering problems
• Available in three variants: Instant for quick use, Thinking for deep analysis, and Pro for maximum accuracy
• Reduces hallucinations by 30% compared to GPT-5.1 Thinking
• Reaches nearly 100% accuracy in analyzing documents up to 256,000 tokens (equivalent to hundreds of pages)
• Improves from 64.2% to 86.3% in understanding graphical interfaces and professional screenshots
• Solves 40.3% of expert-level mathematical problems, compared to 31% for its predecessor
• API price higher than GPT-5.1 but lower than other reference models in the market