OpenAI Releases GPT-5.2 Model Family Amidst Rising Competition

OpenAI has launched GPT-5.2, describing it as its "most powerful model to date," targeting developers and professionals. The release follows an internal "red alert" issued by OpenAI CEO Sam Altman a month prior, acknowledging competitive pressures on ChatGPT and a decline in consumer market share. The new model family aims to reassert OpenAI's leadership position and coincides with the company's tenth anniversary.

GPT-5.2 Model Variations and Capabilities

The GPT-5.2 family comprises three versions: Instant, Thinking, and Pro. Instant is optimized for speed, designed for routine tasks such as information retrieval, writing, and translation. Thinking is tailored for complex structured tasks, demonstrating performance in programming, long document analysis, mathematical calculations, and project planning. Pro represents the highest tier, engineered for critical tasks demanding maximum accuracy and reliability, albeit with slower processing speeds and higher costs.

According to OpenAI, the Pro version is the only model to exceed 90% on the ARC-AGI-1 reasoning benchmark and achieve a perfect 100% score in the AIME 2025 math competition without external tools. All three GPT-5.2 versions are now available within ChatGPT, initially for paying subscribers, with a gradual rollout planned to ensure system stability.

Fidji Simo, CEO of OpenAI's Applications Business, stated that GPT-5.2 was designed to generate economic value. She highlighted significant improvements in spreadsheet creation, presentation development, code writing, image recognition, long-text comprehension, tool utilization, and complex multi-step project handling.

Performance Benchmarks and User Impact

OpenAI reports that ChatGPT Enterprise Edition users save 40-60 minutes daily, with heavy users saving over ten hours weekly. GPT-5.2 aims to extend this value, having set new records in various industry benchmarks.

In the GDPval test, which assesses over 40 professional expertise areas, GPT-5.2 Thinking performed at an expert level. Professional reviewers found that GPT-5.2 Thinking matched or surpassed top industry professionals in 70.9% of cases, including tasks like creating presentations and spreadsheets. The model completed these tasks over 11 times faster than humans, at less than 1% of the cost, though OpenAI did not detail the cost calculation methodology.

For software engineering, GPT-5.2 Thinking achieved a score of 55.6% in the SWE-Bench Pro test, outperforming Claude 4.5 Sonnet and Gemini 3 Pro. It scored 80% in the more fundamental SWE-bench Verified test.

Aidan Clark, OpenAI's Head of Research, noted that enhanced mathematical ability reflects a model's capacity for multi-step logical reasoning, numerical consistency, and error avoidance.

In scientific problems, GPT-5.2 Pro scored 93.2% in the GPQA Diamond test, with GPT-5.2 Thinking achieving 92.4%. Both scores exceeded the previous record held by Gemini 3 Pro. Clark cited a case study where a senior immunology researcher used GPT-5.2 Pro to generate unsolved questions about the immune system, finding the model's questions and explanations "sharper" and superior to other frontier models.

Reliability and Advanced Capabilities

GPT-5.2 also shows progress in reliability. Max Schwarzer, OpenAI's Head of Post-Training, reported a 38% reduction in hallucination rates for GPT-5.2 Thinking compared to GPT-5.1 in factual question-answering benchmarks.

For long-text comprehension, GPT-5.2 Thinking set a new record in the MRCRv2 evaluation, which measures a model's ability to integrate information across extensive documents. When processing real-world tasks requiring information from hundreds of thousands of tokens, GPT-5.2 Thinking's accuracy significantly surpassed GPT-5.1 Thinking. It achieved nearly 100% accuracy in the four-shot MRCRv2 test (up to 256k tokens). This capability allows professionals to process long documents such as reports, contracts, and research papers with greater confidence.

In visual capabilities, GPT-5.2 Thinking improved accuracy by nearly 50% in chart reasoning and software interface understanding. This enables more accurate interpretation of dashboards, product screenshots, and technical diagrams. The model demonstrates a stronger grasp of element positions in images; for example, it can identify major regions and approximate bounding boxes on motherboard images, even with low quality, a significant improvement over GPT-5.1's limited spatial understanding.

OpenAI's new image generation tool remains unreleased. Sam Altman reportedly indicated in an internal memo that image generation would be a future focus, particularly after Google's Nano Banana release. Unconfirmed reports suggest OpenAI plans to release another model in January next year with improved image effects, speed, and "personality."

OpenAI acknowledged areas for improvement, such as addressing over-rejection issues in ChatGPT and enhancing response reliability. The company is also reportedly considering relaxing adult content restrictions for its models.