OpenAI rolls out new GPT-5.2 model

OpenAI has launched GPT-5.2, its latest AI model, on December 11.

The new model is now available to paid ChatGPT users and developers via API.

OpenAI claims GPT-5.2 outperforms previous versions in tasks such as spreadsheet creation, presentation building, coding, and image analysis.

On the GDPval benchmark, which tests knowledge work across 44 occupations, OpenAI said GPT-5.2 matched or exceeded human professionals in 70.9% of cases.

In software engineering tests (SWE-Bench Pro), the model scored 55.6%, up from 50.8% for GPT-5.1.

OpenAI also reported improvements in mathematical reasoning and vision tasks.

The company said GPT-5.2 delivers faster and cheaper outputs for some tasks compared to human experts, though results may vary.

Access to GPT-5.2’s advanced features in ChatGPT requires a Plus, Pro, Business, or Enterprise plan.

🔗 Source: OpenAI

🧠 Food for thought

Implications, context, and why it matters.

OpenAI’s self-reported benchmarks await independent validation

OpenAI says GPT-5.2 leads on GDPval (a cross-occupation knowledge-work benchmark), SWE-bench Pro (a software engineering bug-fixing benchmark), and GPQA Diamond (a graduate-level science question-answering test) 1. These scores need outside checks to separate marketing from gains.
The LMSYS (Large Model Systems Organization) Chatbot Arena runs crowdsourced judging with Elo ratings from head-to-head human votes 2. Results will reveal if GPT-5.2 beats Google’s Gemini 3 (Google’s frontier model), Claude (Anthropic’s assistant), or Llama (Meta’s open-source family) in use.
OpenAI says there is a 38% drop in hallucinations versus GPT-5.1 1. Hallucinations mean fabricated or incorrect outputs presented as facts, so testers need to vet this across tasks before enterprises trust it for mission-critical work.

Migration and optimization opportunities for AI infrastructure providers

GPT-5.2 costs $1.75 per million input tokens for the Thinking tier and $21 per million input tokens for Pro 1. Tokens are units of text the model processes. These prices create demand for services that migrate workloads and calculate ROI on current deployments.
The 400,000-token context window, plus a 128,000 max output 1, enables tasks like processing entire codebases or document sets in one call. This unlocks work for AI consultants and systems integrators to redesign workflows that once needed chunking (splitting content into smaller pieces).
Teams building routing layers like LMSYS’s RouteLLM framework (an open-source router that selects the best model or tier per request) 3 can steer traffic across three GPT-5.2 tiers by task complexity to trim costs while keeping quality 1.