OpenAI Releases gpt-image-2: 2K Resolution Image Generation Now Available in Codex

A New Generation of Image Intelligence

OpenAI released gpt-image-2 on April 21, 2026, positioning it as the most significant advancement in AI image generation since the original gpt-image-1. The model is immediately available through the OpenAI API, the Image Playground, and — critically for developers — directly within Codex, where it can be used to generate visual assets as part of development workflows.

At launch, Sam Altman described the quality jump from gpt-image-1 to gpt-image-2 as equivalent to the leap from GPT-3 to GPT-5. Within hours of the announcement, the model reached #1 across every Image Arena leaderboard, with an unprecedented +242 point lead in the Text-to-Image category.

Reasoning Before Rendering

The most distinctive capability introduced in gpt-image-2 is a native reasoning layer that activates in "Thinking" mode. Rather than immediately rendering pixels, the model researches, plans, and verifies the structure of an image before producing output. In practice, this means the model can:

• Search the web for real-time reference information before generating (e.g., reviewing a product catalog before creating an advertisement)
• Generate multiple coherent variations from a single prompt and compare them
• Self-check outputs for consistency errors before delivering a result
• Produce multi-page documents like comic strips or magazine spreads where recurring characters, layouts, and storylines stay coherent across panels

2K Resolution and Superior Text Rendering

Previous image generation models have historically struggled with rendering legible, correctly-spelled text inside images. gpt-image-2 addresses this as a first-class priority. The model excels at small iconography, UI mockups, product labels, signage, and dense compositional layouts — producing outputs that are difficult to distinguish from professional design work.

Resolution support extends up to 3840×2160 pixels with the outputQuality: "high" parameter, a substantial jump over prior models. Multiple aspect ratios are supported natively, making the model practical for a wide range of production targets including mobile apps, advertisements, presentations, and social media formats.

Multilingual Support

Beyond Latin-script text, gpt-image-2 delivers high-fidelity rendering of Japanese, Korean, Chinese, Hindi, and Bengali — languages where earlier models produced garbled or inaccurate characters. This makes the model viable for international product teams who need localized visual assets generated programmatically.

Developer API Access and Pricing

Developers can access gpt-image-2 through the standard OpenAI API. Pricing is token-based:

• Image input: $8.00 / 1M tokens ($2.00 cached)
• Text input: $5.00 / 1M tokens ($1.25 cached)
• Output: $30.00 / 1M tokens (image), $10.00 / 1M tokens (text)

As a practical reference point, a single high-quality 4K image consumes approximately 13,342 output tokens, costing roughly $0.40. The model is also compatible with the Batch API, which offers a 50% discount on image generation and editing tasks — making large-scale asset generation pipelines significantly more cost-effective.

Available in Codex

For Codex users, gpt-image-2 is available directly within the Codex environment, continuing the platform's expansion into a full creative and development workspace. This follows the image generation capabilities introduced in the April 16 Codex App release (26.415) and gives developers a production-grade image model to work with programmatically alongside their code.