Luma AI’s new image model beats Google and OpenAI on reasoning benchmarks

Luma AI has released Uni-1, an image generation model that outperforms Google’s Nano Banana 2 and OpenAI’s GPT Image 1.5 on key reasoning benchmarks while costing up to 30 percent less at high resolution. Michael Nuñez reports for VentureBeat that the model is part of a significant shift in how AI creates images.

Unlike most older image generators, Uni-1 does not use diffusion, a process that refines random noise into a picture. Instead, it uses autoregressive generation, the same token-by-token method that powers large language models. This means the model reasons through a prompt as it generates the image, rather than separating understanding from creation. This is also how Google’s and ChatGPT’s image generator work today, but different from Midjourney for example.

On the RISEBench reasoning benchmark, Uni-1 scores 0.51 overall, ahead of Nano Banana 2 at 0.50 and GPT Image 1.5 at 0.46. Its advantage is sharpest in logical reasoning, where it scores 0.32 compared to GPT Image’s 0.15. On object detection, it nearly matches Google’s Gemini 3 Pro.

At 2K resolution, a single Uni-1 image costs roughly nine cents via API, compared to about ten cents for Nano Banana 2. Google retains a price advantage at lower resolutions.

Google’s top text-to-image spot remains unchallenged for now, but Uni-1 leads on overall human preference ratings.