XNews

OpenAI GPT-5.2 release and performance claims

OpenAI GPT-5.2 release and performance claims

OpenAI released GPT-5.2 as a work-focused model. - OpenAI released GPT-5.2. - GPT-5.2 designed for “real work,” not demo effects. - Model reportedly surpasses human experts on ~70% of real job tasks. - GPT-5.2 scored a perfect result on AIME 2025. - Model explicitly built for long-horizon (long-cycle) Agents. - Early users report highly creative and extreme new use cases. - A user tested a multimodal comparison case: removed annotation boxes then sent same image to GPT-5.2 and Gemini 3.0. - In that test, Gemini 3.0 outperformed GPT-5.2 on the unannotated image.

Key insights

OpenAI releases GPT-5.2 targeting real-world tasks

OpenAI announced GPT-5.2 as a new model intended for practical, production work rather than demonstration-only use. The release description emphasizes long-horizon agent support and positions the model for sustained task execution in real jobs. The announcement highlights that GPT-5.2 was built with use cases beyond short demos in mind and that developers have already begun experimenting with unconventional applications.

Performance claims: 70% of real tasks and AIME 2025 perfect score

The release materials or reporting state that GPT-5.2 surpasses human experts on approximately 70% of evaluated real-job tasks and achieved a perfect score on AIME 2025. Those specific metrics were presented as part of claims about the model’s frontier performance on practical workloads and benchmarks.

User test challenges multimodal superiority claim versus Gemini 3.0

A user conducted a comparative multimodal test by removing annotation boxes from an image and sending the unannotated image to both GPT-5.2 and Gemini 3.0. In that case, Gemini 3.0 outperformed GPT-5.2, contradicting a claim that GPT-5.2’s multimodal capability exceeded Gemini 3.0 on that example.

Source posts