Reve 2.0 skips the trillion-dollar GPU arms race, lands at #2 on Arena 🎨

Reve released version 2.0 of its AI image model on June 3, debuting at #2 on the Arena text-to-image leaderboard behind OpenAI's GPT Image 2 and ahead of Google's Nano Banana 2. The startup said the system was trained on 10x fewer GPUs than the largest model providers and described it as the best image model not built by a trillion-dollar company.

The model targets a structured "layout" representation in which each object carries a defined location, size, and caption, which the model then reasons over in a thinking trace before rendering at native 4K, equivalent to 16 megapixels. The approach is intended to let users reposition subjects, edit text on signs, or replace backgrounds without regenerating the full image, and the company said API generations run at roughly a fraction of a cent each.

Decrypt's own head-to-head testing found GPT Image 2 holds a slight edge in pure realism, while Reve's outputs retain a filmic, photojournalistic look closer to Midjourney and Flux than the glossier Nano Banana 2. The publication's eight-test review noted that Reve's strength lies in iteration, typography, high-resolution output, and agentic pipelines, while general-purpose bundles from Google and OpenAI may suit users who need more than image generation.