OpenAI has unveiled GPT-5.2, the latest in its GPT-5 series, aimed at professional knowledge work and complex agentic tasks.

The rollout starts immediately, with paid users gaining early access, marking a significant step in AI capabilities and safety improvements.

GPT-5.2 is not a single model but a series: GPT-5.2 Instant, GPT-5.2 Thinking, and GPT-5.2 Pro. OpenAI claims this release improves performance in math, science, coding, imaging, agentic tasks, and general accuracy, making it the most capable model series yet for professional work.

According to OpenAI, GPT-5.2 produces fewer hallucinations than GPT-5.1. For example, GPT-5.2 Thinking has an average hallucination rate of 10.9%, down from 12.7% in GPT-5.1 Thinking. When combined with web access, this drops further to 5.8%, improving reliability for real-world tasks.

Enhanced abilities across professional workflows

GPT-5.2 demonstrates notable improvements in spreadsheets, presentations, and agentic tasks, performing at or above human expert levels in GDPval, a benchmark spanning 44 occupations. Outputs are generated up to 11x faster and at less than 1% the cost of human experts.

GPT-5.2 sets a new standard in coding performance, achieving 55.6% on SWE-Bench Pro, which evaluates software engineering across multiple languages. The model also excels in interactive coding, code review, and bug detection, according to internal and third-party evaluations.

Scientific, math, and vision benchmarks

The model scored higher on scientific reasoning, advanced mathematics, abstract reasoning, and computer vision tasks compared to previous GPT-5 iterations. Specialized tools in GPT-5.2 enable complex, multi-step task execution more effectively than ever before.

Focus on safety and mental health

OpenAI has aimed to improve GPT-5.2’s responses in conversations involving mental health issues, suicide risks, and emotional reliance. Internal tests suggest the model produces fewer undesirable outputs than GPT-5.1, a response to previous safety criticisms.

This update follows lawsuits and criticisms that previous models encouraged harmful behavior in vulnerable users. OpenAI says GPT-5.2 contains targeted interventions to reduce risky outputs in sensitive contexts.

Availability and Access

GPT-5.2 is available immediately to paid Plus, Pro, Go, Business, and Enterprise users, with a gradual rollout planned to ensure stability. Legacy GPT-5.1 models will remain available for three months.

Developers can access GPT-5.2 via API, enabling integration for professional, research, and agentic applications.