OpenAI has announced ChatGPT Images 2.0 powered by the updated gpt-image-2 model. The tool immediately claimed the top spot across all image generation and editing categories in the LM Arena benchmark, outperforming its closest competitor by 242 Elo points. Such a high result is attributed to high-quality multilingual text rendering and strict prompt adherence.
Thinking Before Drawing
The key change is the integration of reasoning algorithms used in the O-series model family. Before starting generation, the neural network analyzes the prompt, plans the composition structure, and verifies object counts and prompt constraints. Paid ChatGPT subscribers (Plus, Pro, Business) have access to this feature as thinking mode. In the free version, the generator operates in standard mode without the pre-planning stage.
For developers using the API, a thinking parameter is available with three levels of detail: low, medium, and high. Additionally, during reasoning, the model can perform real-time web searches, which is especially useful for creating infographics and diagrams with accurate factual data.
Feature Comparison: gpt-image-1 vs gpt-image-2
| Parameter | gpt-image-1 | gpt-image-2 |
|---|---|---|
| Maximum resolution | 1024 px | 2000 px on the long side |
| Supported aspect ratios | 1:1, 3:2, 2:3 | 1:1, 3:2, 2:3, 16:9, 9:16, 3:1, 1:3 |
| Number of images per prompt | 1 | up to 10 (in a consistent style) |
| Text rendering | basic English only with errors | multilingual (including Cyrillic, CJK, and Indic scripts) |
| Reasoning mechanism | none | supported (thinking parameter) |
| Web search | no | available in thinking mode |
| Knowledge cutoff | — | December 2025 |
Pricing, Availability, and Sunsetting Legacy Versions
API pricing consists of several components: $5 per million input and $10 per million output text tokens, plus $8 per million input and $30 per million output image tokens. A standard 1024×1024 image costs approximately $0.21, which is 60% higher than the first generation due to reasoning computation and an expanded canvas size.
The model is integrated into the OpenAI Codex development environment, while third-party platforms such as Figma, Canva, and Adobe Firefly have already announced support. Additionally, OpenAI announced plans to fully deprecate legacy DALL-E 2 and DALL-E 3 on May 12, 2026.
Current Limitations
Despite the technological leap forward, the model still has weaknesses: close-up human portraits may still show defects, geometrically precise brand logos remain challenging, and long coherent text exceeding a few hundred characters begins to lose structure. Furthermore, stylistic consistency is maintained only within a single batch generation and resets across different conversations.
Comments
to leave a comment.
No comments yet.