OpenAI releases GPT-5.2 after a "code red"

Background

Reuters described the launch as following a "code red" push at OpenAI to counter Google's Gemini 3 ^[2]. That framing placed the release in a competitive context: the model read as an answer to a rival rather than a routine update. A "code red" in this usage means an internal, high-priority effort that pulls resources toward a single goal.

What was launched

On 11 December 2025 OpenAI launched GPT-5.2 in three versions: Instant, Thinking and Pro ^[1]. The model was included in ChatGPT and in Microsoft Copilot ^[1], so it was part of both OpenAI's own chat product and Microsoft's assistant. Within ChatGPT, the three variants began rolling out to paid plans that day, while the API versions were available to all developers immediately ^[4]. Paying ChatGPT users therefore met the model inside the app, while developers could call it through the API from day one ^[4].

Professional work

OpenAI said the model was better at professional tasks such as spreadsheets, financial models and presentations ^[1][3]. The evidence it offered was GDPval, an evaluation that compares model output with the work of professionals across different occupations. OpenAI said GPT-5.2 Thinking beat or tied top industry professionals on 70.9% of GDPval comparisons across 44 occupations, and that it was the company's first model to perform at or above human expert level on that evaluation ^[4]. OpenAI also said the model produced GDPval outputs more than 11 times faster than expert professionals and at under 1% of their cost ^[4].

Benchmark scores

OpenAI reported 55.6% on SWE-Bench Pro and 80.0% on SWE-bench Verified, two benchmarks built around software engineering tasks ^[4]. On GPQA Diamond, a set of difficult science questions, it reported 92.4% without tools; on AIME 2025, a high-school mathematics competition, 100% without tools ^[4]. "Without tools" means the model was not given external aids for those runs, so the figures describe the model's own output rather than output produced with calculators or code execution.

How to read the numbers

All of the performance and cost figures above were reported by OpenAI about its own model and its own evaluations ^[4]. The human-comparison claim rests on GDPval, where 70.9% of comparisons were counted as a win or a tie for the model ^[4]; the speed and cost comparison is stated relative to expert professionals ^[4].

What it signalled

CNBC reported that Sam Altman expected OpenAI to exit "code red" by January 2026 ^[3], which put an approximate timetable on the internal push that preceded the launch.