Forge
Skip to main content

OpenAI Introduces GPT-5.2: An incremental upgrade or something more?

Last updated: December 12, 2025
OpenAI has launched GPT-5.2, a new frontier model aimed at advancing professional workflows, long-running agents, and complex knowledge work. The company positions GPT-5.2 as its most capable model series yet, surpassing earlier versions in intelligence, speed, and reliability across a wide range of applications.


Core Advancements

GPT-5.2 introduces improvements in reasoning, tool use, document handling, vision, and agentic task execution. OpenAI reports significant performance gains in tasks involving spreadsheets, presentations, software development, and long-horizon workflows, making the model more useful for enterprise teams and power users who handle large, multi-step projects.

Benchmarks and Professional Performance

GDPval, a benchmark that captures real professional tasks across 44 occupations, shows GPT-5.2 Thinking outperforming or tying industry professionals in over 70 percent of evaluated tasks. The model not only matches expert-level outputs more often but generates them at dramatically lower cost and faster speed. Internal tests also show notable gains in financial modeling tasks common in investment banking.

Coding Improvements

GPT-5.2 Thinking sets a new high score on SWE-Bench Pro, reflecting stronger software engineering abilities across multiple languages. Early testers observed substantial improvements in debugging, refactoring, and interactive coding sessions, including more advanced front-end and 3D interface work. Companies like Windsurf, JetBrains, and Augment Code report that the jump in capability feels larger than a minor version change.

Factuality and Reliability

GPT-5.2 reduces hallucination rates when compared with GPT-5.1. Across a broad sample of real de-identified user queries, erroneous responses decreased by roughly a third. This shift strengthens the model’s role in research, analysis, and decision support, although OpenAI continues to caution that critical outputs should be reviewed.

Long Context Understanding

The model reaches new highs in long-context reasoning, particularly on OpenAI’s MRCRv2 evaluation. GPT-5.2 Thinking approaches perfect accuracy in scenarios involving multiple embedded tasks spread across hundreds of thousands of tokens. This makes it more effective for reviewing long reports, legal documents, research papers, or multi-file technical projects. The new compact mode can extend effective context even further for tool-heavy workflows.


Vision Upgrades

GPT-5.2 improves accuracy on chart interpretation, scientific figure reasoning, and software interface understanding. The model produces more coherent spatial interpretations and bounding-box predictions than earlier versions, strengthening its utility for engineers, analysts, designers, and support teams who rely on visual information.

Tool Calling Advances

On τ2-bench Telecom, GPT-5.2 achieves near-perfect tool-calling performance at 98.7 percent. Improvements are also visible at low reasoning effort, supporting latency-sensitive use cases. The model can coordinate long sequences of actions across tools and agents more reliably, as shown in scenarios like resolving complex customer-support cases involving rebooking, compensation, and special accommodations.

Science and Math Capabilities

GPT-5.2 Pro and GPT-5.2 Thinking push further into scientific reasoning and quantitative problem solving. Scores reach 93 percent on GPQA Diamond and set new state-of-the-art performance on FrontierMath Tier 1–3. OpenAI highlights early research where GPT-5.2 contributed to a proposed proof in statistical learning theory, later vetted by experts.

User Experience in ChatGPT

ChatGPT users will encounter three variants: Instant, Thinking, and Pro. Instant focuses on speed and clarity for everyday tasks. Thinking offers structured, deeper reasoning for complex projects. Pro targets the highest accuracy for challenging domains. Early testers describe GPT-5.2 as more polished, more organized, and more stable in long dialogues and project-level tasks.

Safety Updates

GPT-5.2 incorporates the safety-completion techniques introduced in GPT-5, with additional tuning for mental health, emotional reliance, and sensitive topics. Models perform substantially better on internal safety benchmarks measuring responses to self-harm and distress-related prompts. OpenAI is also rolling out an age-prediction system to automatically apply content protections for minors.

Availability and Pricing

GPT-5.2 is rolling out now for Plus, Pro, Business, Enterprise, and Go customers, with GPT-5.1 remaining available for a three-month sunset period within ChatGPT. In the API, GPT-5.2 Thinking is listed as gpt-5.2 and GPT-5.2 Instant as gpt-5.2-chat-latest. Pricing is higher than GPT-5.1 but still below other frontier models. GPT-5.2 Pro is available for tasks requiring maximum reasoning quality.

The Bigger Picture

OpenAI frames GPT-5.2 as part of a continuous upgrade cycle focused on intelligence, reliability, tool coordination, and safety. While the release brings clear improvements, the company notes ongoing work on issues like over-refusal and the need for even greater stability. With stronger long-context performance, better factuality, and more capable multi-step reasoning, GPT-5.2 marks another step toward AI systems that can act as reliable partners in complex, real-world professional workflows.
Tags: ML News