Thinks First, Then Creates: ChatGPT Images 2.0 Marks a Breakthrough Step by OpenAI

OpenAI has officially unveiled ChatGPT Images 2.0 (gpt-image-2), a new generation image model designed as a direct response to Google's competing solution, Gemini Nano Banana 2. Previously developed under the codename "duct tape," the system introduces major upgrades — built-in reasoning capabilities, significantly improved text rendering, and enhanced multilingual support.

Reasoning, baked into the image generator

The most notable innovation is the integration of "O-series" reasoning directly into the image generator. Unlike traditional models that act as a "black box," the Thinking version operates more like an agent. It can:

• analyze data,
• browse the web in real time,
• process uploaded files (such as PowerPoint presentations),
• and plan the structure of an image before rendering it.

As a result, the model goes beyond simply "drawing" and can produce well-structured, logical outputs such as:

• complex infographics and maps with accurate data representation and clear legends,
• educational materials spanning multiple pages while maintaining visual and conceptual consistency,
• interior design concepts and visual systems, including floor plans, color palettes, and material lists.

Text rendering: finally fixed

The model also addresses one of the biggest weaknesses of earlier image generators: incorrect text rendering. OpenAI describes this improvement as a "step change."

Images 2.0 can accurately generate text even in dense layouts like restaurant menus, magazine covers, or user interfaces. It has also become effectively multilingual, with much stronger support for non-Latin scripts such as Japanese, Chinese, Korean, Hindi, and Bengali. Text in these languages is not just translated but naturally integrated into the visual design.

Under the hood

OpenAI has completely reworked the model's architecture and has not disclosed whether it is diffusion-based or autoregressive. However, several technical capabilities are known:

• image generation up to 2K resolution in ChatGPT and up to 4K via the API (beta),
• support for a wide range of aspect ratios, from 3:1 panoramas to 1:3 vertical formats,
• the ability to generate up to 8 consistent images from a single prompt (useful for comics or storyboards),
• knowledge updated through December 2025.

Access tiers

Access to the model is divided into tiers:

• Free and Codex users get access to Images 2.0 Instant — faster generation, improved instruction following, better text handling.
• Plus, Pro, and Business users can use the Thinking model, which includes tools, web browsing, and multi-image generation.
• Pro users additionally gain access to ImageGen Pro for the most advanced results.

API and pricing

For developers, gpt-image-2 is available via Microsoft Foundry and API access, with pricing set at:

• $8.00 per million input tokens,
• $2.00 per million cached input tokens,
• $30.00 per million output tokens — which is $2 cheaper than the previous GPT-Image-1.5 model.

Safety and disinformation

OpenAI emphasizes a strong focus on safety, especially given the rise of disinformation campaigns and deepfakes. Images 2.0 includes multi-layered safeguards such as watermarking and advanced content filters. The company also maintains strict policies against election interference and the creation of misleading political content.

Bottom line

Images 2.0 isn't just another bump on the quality ladder — it's the first time a major image generator plans like an agent before it renders. Combined with the text-rendering step change and proper multilingual support, it closes the biggest gaps that still forced designers and educators back to manual tools. The Google vs. OpenAI race on generative imagery just got significantly more interesting.