OpenAI’s New Image Model Hazelnut: Early Insights & Features

The generative AI landscape is on the edge of another major shift. Over the last 48 hours, unannounced image generation architectures attributed to OpenAI have been identified on public testing arenas, offering a compelling first look at the company’s next-generation visual capabilities.

Currently undergoing blind benchmarking under the codename Hazelnut, this model appear to be the foundation for the highly anticipated GPT Image 2. Early performance data suggests OpenAI is moving beyond standard image synthesis, targeting high-fidelity technical rendering and enhanced world knowledge to challenge the market's top contenders.

Here is a summary of the insights and technical specifications known so far.

1. The Model Hierarchy: Hazelnut & Chestnut

Observations from platforms like Design Arena and LM Arena indicate a two-tier strategy, likely designed to balance performance with efficiency:

• Hazelnut: This appears to be the flagship model and the likely core of GPT Image 2. Performance metrics suggest it is optimized for maximum fidelity, complex prompt adherence, and high-resolution output.

• Chestnut: Likely a lightweight, optimized variant (potentially a "Mini" version of the architecture), engineered for faster generation speeds and lower latency.

2. Technical Breakthroughs: Text & Code Rendering

Prompt: Create a screenshot of a computer screen displaying Google Docs open in a Chrome browser. Inside the Google Docs document, the following essay should be fully written:“The Power of Small Daily Habits

The most distinct improvement noted in early testing is the capability to integrate complex text and technical structures - a feature expected to be a selling point for GPT Image 2. Emerging test results show:

• Code Generation: The model can render accurate, legible code snippets within an image, making it a viable tool for generating UI mockups, IDE screenshots, and technical documentation.

• Data Visualization: Users have reported high success rates with mathematical formulas, flowchart labels, and diagrammatic elements, which are rendered without the text distortion common in previous generations.

3. Competitive Benchmarking: Rivaling "Nano Banana Pro"

Prompt: a split photo of berlin from 1938 and may 8th, 1945 in the same spot, demosntrating the absolute destruction of the city. color photograph historical photo, no labels.

The new architectures are being actively compared against Google’s current high-performance model, referred to in testing circles as Nano Banana Pro. Key performance indicators include:

• World Knowledge: The GPT Image 2 candidates demonstrate a sophisticated understanding of specific objects, cultural context, and nuance, rapidly closing the gap with Google’s existing standards.

• Photorealism & Likeness: Early outputs show a marked improvement in generating realistic human subjects, including public figures. This suggests a refinement in safety tuning to reduce refusal triggers while minimizing the "uncanny valley" effect.

• Current Limitations: While the advancements are significant, some comparative analyses note that skin textures can occasionally appear "plasticky" relative to the competitor, indicating that final fine-tuning may still be underway.

The Battle for Supremacy Heats Up.

The sudden appearance of Hazelnut and Chestnut makes one thing clear: OpenAI is not conceding the image generation domain to Google’s Nano Banana Pro. With the potential arrival of GPT Image 2, the competition is shifting from simple image generation to high-utility, multimodal intelligence. As these models exit the testing phase, the industry is about to see whether OpenAI can effectively reclaim the throne of high-fidelity image synthesis.