OpenAI's next-generation image generation model appears to have surfaced, with blind tests already underway on Design Arena and LM Arena. December 9, 2025, San Francisco — Several independent testers discovered today that OpenAI is conducting small-scale blind tests on two new image generation models codenamed "Chestnut" and "Hazelnut" on the AI evaluation platforms Design Arena and LM Arena.

QQ20251210-141417.png

This marks the most significant development signal from OpenAI in the text-to-image field since the official release of gpt-image-1 in May this year. According to publicly available blind test samples and scoring data, the new model shows significant improvements in multiple key dimensions:

  • Compared to Google's latest Nano Banana Pro, its understanding of world knowledge has become almost comparable;
  • It can generate celebrity-style selfies with near-photographic realism, with facial details, facial proportions, and lighting treatment significantly better than gpt-image-1;
  • It excels at embedding readable code within images, accurately rendering complex code snippets, flowchart labels, and mathematical formulas, virtually eliminating common issues such as text distortion and hallucination;
  • The overall image structural integrity, color accuracy, and style consistency are approaching the current industry's highest standards.

Currently, both models are participating in the rankings anonymously. Chestnut is believed to be a lightweight version (corresponding to the future "Image-2-mini"), while Hazelnut is likely the flagship version (corresponding to "Image-2").

QQ20251210-141358.png

Industry insiders analyze that this blind test is typically a routine process 1-3 weeks before a major model launch by OpenAI. Combined with previous leaked roadmaps, the new image model is highly likely to be released alongside the rumored GPT-5.2, with the official announcement possibly coming as early as this week or next week.

If confirmed, this would be the largest leap in image capabilities for OpenAI since DALL·E3, 14 months ago, and would also allow it to regain the initiative in direct competition with rivals such as Google, Midjourney, and Flux. The OpenAI official has not yet responded to this.