Within 12 hours of launch, GPT Image 2 topped the Arena text-to-image leaderboard with a score of 1512, leading the runner-up Nano Banana 2 by 241 points -- the biggest gap in Image Arena history. It is not only OpenAI's counterattack after half a year of pressure from Google; more importantly, it pushes image generation from the "rendering era" into the "reasoning era".

What Is GPT Image 2

GPT Image 2 is an image generation model rebuilt from scratch by OpenAI, which officially calls it "the GPT of images". Its biggest difference from all previous image models: native thinking ability -- it plans before drawing, checks its own work after drawing, and can even search the web for reference material when needed.

GPT Image 2 topped all three Arena sub-leaderboards; the official wording is "a clean sweep"

Core Capabilities

99% Text Rendering Accuracy

Text rendering has always been a hard problem for AI image generation. GPT Image 2 pushes text rendering accuracy to 99% (the previous generation, Nano Banana Pro, was at 94%). Its text rendering sub-score rose 316 points over the previous generation.

Multiple Styles from One Prompt

The old pain of AI image generation was uniform output. GPT Image 2 can produce multiple images in different styles from a single prompt -- different compositions, color schemes, and information density -- a big step up in variant controllability.

Linear algebra explainer images generated from the same prompt, in four completely different compositions and styles

Multi-Image Editing

Beyond generation, GPT Image 2 also topped the single-image and multi-image editing sub-leaderboards. It can make precise modifications directly on top of the original image, completing the edit while preserving the original's details.

Thinking Mode

With Thinking mode enabled, the model first produces a reasoning plan and then starts drawing. This mode is especially good at complex scenes -- comic page generation, for example, where it keeps character consistency and coherent multi-panel storytelling.

A comic page generated in Thinking mode, with character consistency and coherent multi-panel storytelling

How to Use It

In ChatGPT

ChatGPT Plus subscribers ($20/month) can use it directly; image generation is already included in the plan.

API

from openai import OpenAI

client = OpenAI()

response = client.images.generate(
    model="gpt-image-2",
    prompt="A realistic candid photograph, coastline, overcast sky, a vintage car parked at the roadside, film texture",
    n=1,
    size="1024x1024"
)

API pricing: $0.21 per image for high-quality output.

Third-Party Integrations

On launch day, Figma, Canva, Adobe Firefly, fal, and Hermes Agent had all completed integrations, so you can call GPT Image 2 directly inside these tools.

Note: DALL-E 2 and DALL-E 3 will be officially retired on May 12.

Real-World Comparison

Arena blind-test data: GPT Image 2's win rate is 93% -- across 100 paired blind image selections, people picked GPT Image 2's result 93 times.

Limitations

• Cannot faithfully reproduce brand logos (in ZDNet's hands-on test, even its own logo came out crooked)
• Nano Banana 2 still has the edge in photorealistic portraits and multi-reference-image consistency
• Complex scenes occasionally show detail errors