News

OpenAI Launches GPT-5.2 in Three Versions After "Code Red" Push to Counter Google's Gemini 3

• By John K. Waters
• 12/16/2025

OpenAI has released GPT-5.2, the latest version of its AI model family for ChatGPT, following a reported "code red" directive from CEO Sam Altman earlier this month. The directive was issued in response to competitive pressure from Google's Gemini 3 model, which recently topped multiple AI benchmarks.

GPT-5.2 is available in three versions: Instant, Thinking, and Pro. OpenAI says each variant is tailored to different user needs. The Instant model handles lightweight tasks such as writing and translation, while Thinking is intended for more complex reasoning and problem-solving. The Pro model is designed for users requiring high-accuracy performance in professional and technical domains.

The model features a 400,000-token context window and a knowledge cutoff date of August 31, 2025. Paid ChatGPT subscribers began receiving access on Thursday, with developers able to access GPT-5.2 via the OpenAI API. The company priced API usage at $1.75 per million input tokens, a 40 percent increase over the previous GPT-5.1 model. OpenAI said GPT-5.1 would remain available for three months under a legacy model option.

OpenAI's release comes after a turbulent period for the company. In early December, Altman reportedly issued a "code red" alert within the organization, halting other projects—including advertising plans—to prioritize improvements to the ChatGPT experience. The move followed Gemini 3's strong performance across a variety of AI benchmarks and its growing user base.

Google's Gemini app currently reports more than 650 million monthly active users, while OpenAI says ChatGPT serves around 800 million weekly active users. OpenAI has also committed to infrastructure investments totaling $1.4 trillion over the next several years.

GPT-5.2 marks the company's third major model update since August. GPT-5 introduced a new system for routing user prompts between instant and simulated reasoning modes. The 5.1 release in November added preset personality options and improved conversational quality. Some users, however, had criticized earlier 5.x versions for feeling cold and robotic.

Although OpenAI did not publish head-to-head comparisons with Gemini 3 on its public-facing materials, the company did share some benchmark data during the press briefing. GPT-5.2 Thinking reportedly scored 55.6 percent on SWE-Bench Pro, a benchmark for software engineering tasks, ahead of Gemini 3 Pro's 43.3 percent and Claude Opus 4.5's 52.0 percent.

On the GPQA Diamond science benchmark, GPT-5.2 achieved 92.4 percent accuracy, slightly higher than Gemini 3 Pro's 91.9 percent.

The company said GPT-5.2 Thinking outperforms or matches human professionals on 70.9 percent of tasks in its new GDPval benchmark, which measures AI performance across 44 occupations. It claims the model completes these tasks more than 11 times faster and at less than 1 percent of the cost of a human expert.

Although the results appear promising, industry observers remain cautious. Independent validation of benchmarks takes time, and experts continue to note that many corporate performance claims are challenging to verify. The science of objectively measuring AI reasoning remains underdeveloped relative to marketing narratives.

For now, users of ChatGPT can expect incremental improvements in areas such as coding and productivity tasks, as the pace of AI model development continues to accelerate.