OpenAI, the developer of ChatGPT that declared a "code red" in response to Google's pursuit, unveiled its cutting-edge artificial intelligence (AI) model, GPT-5.2. As Google's Gemini 3 rapidly gained users with performance surpassing GPT-5, OpenAI, sensing a crisis, responded with a new version just one month after releasing the previous version.
OpenAI said on the 11th (local time) that it will release GPT-5.2, designed for professional knowledge work. OpenAI said GPT-5.2 is faster and more accurate in information search, writing, and translation, and improved its ability to mimic human reasoning, boosting its capacity to handle complex, long-form tasks such as math and programming.
GPT-5.2 consists of three modes in total, adding a Pro mode to the transfer instant answer (Instant) and thinking (Thinking) modes.
The instant answer mode is optimized for quick responses for everyday tasks and learning, and the thinking mode is suitable for coding, mathematical problem-solving, and summarizing long documents. OpenAI said the Pro mode is suitable for solving difficult problems that require long working hours.
OpenAI emphasized that the latest model strengthens reasoning and coding capabilities and reduces hallucinations, and that it received top scores on various benchmarks measuring AI performance. Fidji Simo, OpenAI applications institutional sector CEO, said, "GPT-5.2 performs best in complex multi-step project execution, such as creating spreadsheets, organizing presentations, writing code, image recognition, long-term context understanding, and tool use."
In the GDPval evaluation, which assesses job performance across 44 professional industry occupations, the GPT-5.2 thinking mode scored 70.9% and the Pro mode 74.1%, and was found to be on par with or better than human experts. Compared with the previous version, GPT-5 (38.8%), performance improved significantly.
On the SWE benchmark, which evaluates software engineering capabilities, it scored 80%, surpassing Gemini 3 Pro's 76.2%. It also rose to nearly the same level as Anthropic's Claude Opus 4.5 (80.9%), known for its strong coding ability. It also outperformed Google's latest AI model on HLE, a benchmark for doctoral-level reasoning ability dubbed "the last exam for humans."
In evaluations conducted without separate tools, relying only on the model's own capabilities, GPT-5.2 Pro scored 36.6%, slightly lower than Gemini 3 Pro's 37.5%. However, in evaluations using tools such as search, it scored 50%, higher than Gemini 3 Pro (45.8%).
Hallucinations were also greatly reduced. In internal tests, the error rate of GPT-5.2's thinking mode was 6.2%, about a 30% decrease compared with its predecessor, GPT-5.1's thinking mode (8.8%).
Its ability to grasp the context of long texts, recognize images and photos including graphs, and solve multi-step tasks also improved.
GPT-5.2 will be rolled out sequentially starting today to paid plan subscribers. Developers can also use it through the API. With the release of the new version, the previous version, GPT-5.1, will be offered for the next three months and then discontinued.
OpenAI initially planned to release GPT-5.2 at the end of the month, but is said to have moved up the launch after seeing Gemini 3 Pro attract users on the strength of its powerful benchmark report card.
OpenAI CEO Sam Altman issued a code red internally earlier this month and directed staff to set aside other work and focus on ChatGPT performance and improvements.
The Wall Street Journal (WSJ) reported that OpenAI plans to lift the alert after unveiling another new model in January next year.