OpenAI Unveils GPT-5.2 for Professional Work, Long-Context Agents

OpenAI releases GPT‑5.2 for professional work and long-running agents

OpenAI published GPT‑5.2 on December 11, 2025, positioning the series as a step forward for professional knowledge work, long-context reasoning, tool-driven agents, and multimodal perception. The rollout began that day in ChatGPT for paid plans and is available in the API for developers. The release highlights improvements across coding, spreadsheets and slides, vision, tool calling, and long-document workflows.

What’s new and notable

• Benchmark gains: GPT‑5.2 Thinking posts large improvements on several public and internal evaluations. Notable wins include a 70.9% score on GDPval knowledge-work comparisons (beating or tying industry professionals), 55.6% on SWE‑Bench Pro and 80.0% on SWE‑bench Verified for coding, and strong results on academic reasoning such as 92.4% on GPQA Diamond.
• Long-context capability: GPT‑5.2 achieves near‑100% accuracy on the 4‑needle MRCR variant out to 256k tokens, enabling coherent analysis across extremely long documents and multi-file projects. The model also supports the Responses /compact endpoint to extend effective context for tool-heavy, long-running workflows — see the API docs: https://platform.openai.com/docs/api-reference/responses/compact.
• Tool calling and agents: The model attains 98.7% on Tau2‑bench Telecom for reliable tool use in multi‑turn tasks, and shows improvements when reasoning.effort is set to 'none' for latency‑sensitive scenarios. These gains matter for orchestrating end‑to‑end agent workflows such as multi‑step customer support, data retrieval, and automated case resolution.
• Vision and charts: Vision performance improves substantially on chart reasoning and UI understanding (e.g., CharXiv and ScreenSpot‑Pro), with error rates roughly halved versus prior models. The model demonstrates better spatial understanding of image elements and improved chart interpretation when paired with tools like Python where applicable.
• Coding and front‑end work: GPT‑5.2 is stronger at debugging, refactoring, and shipping fixes across multiple languages, with particular gains reported for front‑end and complex UI tasks, including 3D elements. The model is presented as more capable for interactive coding, code reviews, and automated bug finding.

Factuality, science, and math

GPT‑5.2 Thinking shows fewer hallucinations than GPT‑5.1—responses with detectable errors were around 30% relatively less common in a de‑identified ChatGPT query set. In scientific and mathematical domains the release cites high marks (for example, 100% on AIME 2025 and improved performance on FrontierMath). The announcement also references collaborative research where GPT‑5.2 Pro assisted in proposing a proof that was verified by researchers.

Safety and sensitive content

The release builds on prior safe-completion work and reports targeted improvements in responses to prompts indicating suicide, self‑harm, mental‑health distress, and emotional reliance. OpenAI notes changes to apply content protections for under‑18 users are being rolled out, and points readers to the updated system card for more details: https://openai.com/index/gpt-5-system-card-update-gpt-5-2/.

Availability, pricing, and developer access

• ChatGPT rollout began December 11, 2025, starting with paid plans (Plus, Pro, Go, Business, Enterprise). GPT‑5.1 remains available to paid users for three months as a legacy model before sunsetting.
• API names and mappings:
  • ChatGPT‑5.2 Instant → gpt-5.2-chat-latest
  • ChatGPT‑5.2 Thinking → gpt-5.2
  • ChatGPT‑5.2 Pro → gpt-5.2-pro
• Pricing (per million tokens):
  • gpt-5.2 / gpt-5.2-chat-latest: Input $1.75; cached input $0.175; Output $14
  • gpt-5.2-pro: Input $21; Output $168
  • For comparison, gpt-5.1 input and output are listed at $1.25 and $10 respectively. Cached inputs receive a 90% discount where applicable.

Developers can set higher reasoning effort levels (xhigh) for GPT‑5.2 Pro and Thinking in the API for tasks prioritizing quality.

Infrastructure and partners

OpenAI credits partnerships with NVIDIA and Microsoft for the training and deployment infrastructure, citing Azure data centers and NVIDIA H100, H200, and GB200‑NVL72 GPUs as part of the compute stack.

For full technical detail, benchmark tables, and the original announcement, see the source on OpenAI’s site: https://openai.com/index/introducing-gpt-5-2/