Robots Atlas>ROBOTS ATLAS
ChatGPT Images 2.0
AI Models›GPT

ChatGPT Images 2.0

gpt-image-2 · Family: GPT
OpenAI's image generation and editing model with built-in O-series reasoning, up to 2K resolution, non-Latin text rendering, and batch generation of up to 8 images per prompt. Available via API as gpt-image-2.
✓ Active✓ Public accessImage generationReasoning modelMultimodal📁 GPT
Release date
21 April 2026
🏢OpenAIProducer
Access:APIHostedDeployment:☁ Cloud

Overview

GPT Image 2 (API name: gpt-image-2) is OpenAI's image generation and editing model, released on 21 April 2026 under the ChatGPT Images 2.0 brand. Unlike the earlier diffusion models DALL·E 2 and DALL·E 3, GPT Image models are autoregressive, and GPT Image 2 adds a reasoning layer based on the O-series model family.

Architecture and reasoning

Before generating an image, the model analyzes the prompt, plans composition and element layout, models spatial relationships, optionally searches the web for current information, and verifies its own output after generation. Earlier diffusion models operated reactively, without an explicit planning step.

Operating modes

  • Instant mode: available free of charge to all ChatGPT users, without a reasoning step.
  • Thinking mode: available on ChatGPT Plus, Pro, Business and Enterprise plans; includes web search, generation of up to 8 coherent images from a single prompt, and output verification.

Generation parameters

  • Resolutions up to 2K (experimentally above 2560×1440 px).
  • Aspect ratios from 3:1 to 1:3.
  • Output formats: PNG, JPEG, WebP.
  • Transparent background is not supported in Responses API mode.

Text rendering

The model significantly improves text rendering within images and supports both Latin and non-Latin scripts (including Japanese, Korean, Chinese, Hindi and Bengali). OpenAI declares a text rendering accuracy of ~99% — a manufacturer figure that has not been independently audited.

Availability and lifecycle

  • Available in ChatGPT (web, mobile, desktop), the Codex environment, and via the OpenAI API.
  • Production snapshot: gpt-image-2-2026-04-21; the chatgpt-image-latest alias points to the current default image model.
  • Replaces DALL·E 2, DALL·E 3 and gpt-image-1.5; deprecation of the previous models is scheduled for 12 May 2026.
Classification
Image generationReasoning modelMultimodal
Family: GPT
Applications
Content generation
Access & deployment
APIHosted
Cloud
Weights: Closed
Key parameters
📥 Input: text, image
Platforms
OpenAI API

Technical specification

Knowledge cutoff
31 Dec 2025
Knowledge boundary
License
Proprietary / Commercial
Hardware requirements
Not applicable — the model is available exclusively via the OpenAI API and ChatGPT (closed cloud). Self-hosting and weight downloads are not supported.
Modalities
⬇ Input
textimage
⬆ Output
image

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Planning
Forming and executing action plans for complex tasks.
Category: planning
Application domains
Content generation

Benchmark results

5 benchmarks
Image Arena — Text-to-Image
Elo score · Evaluated in medium quality mode. Leads the second-ranked model by +242 Elo points over model #2 (Nano Banana 2 / Google Gemini 3.1 Flash Image, score 1271). The largest recorded gap between #1 and #2 in the history of the Image Arena platform.
1512punkty Elo
📅 19 Apr 2026📄 Arena.ai / Image Arena leaderboard (crowdsourced, blind human preference voting)
Arena.ai is an independent crowdsourcing platform — results may change as new votes are submitted. The score refers to medium quality, not high quality.
Image Arena — Single-Image Edit
Elo score · Ranked first. Margin of +125 points over model no. 2 (Nano Banana Pro).
1513punkty Elo
📅 21 Apr 2026📄 Arena.ai / Image Arena leaderboard
Blind human preference voting. Results may evolve over time.
Image Arena — Multi-Image Edit
Elo score · Ranked first. Margin of +90 points over model no. 2 (Nano Banana 2).
1464punkty Elo
📅 21 Apr 2026📄 Arena.ai / Image Arena leaderboard
Blind human preference voting.
Image Arena — Text Rendering (sub-kategoria)
Elo improvement vs GPT-Image-1.5 High Fidelity · Largest gain in the text rendering category among all sub-categories. GPT Image 2 ranked #1 in all 7 Text-to-Image sub-categories.
+316punkty Elo (poprawa względem poprzednika)
📅 21 Apr 2026📄 Arena.ai / Image Arena category breakdown; raport officechai.com
Data from media coverage of the Arena sub-category review. For reference only.
Wewnętrzny benchmark text rendering (OpenAI)
Text accuracy · Accuracy of text rendering as declared by OpenAI. The previous model, gpt-image-1.5, achieved 90–95%. The methodology has not been publicly disclosed.
~99%procent
📅 21 Apr 2026📄 OpenAI press release przy premierze 21.04.2026
Manufacturer's declaration — not independently audited.

Pricing

Technical architecture

Core Architecture
AGAR Generation
Model Form
NMNative Multimodal
Training Techniques
RLRLHFITInstruction TuningCOCoT

Deployment and security

☁ Available on platforms
☁OpenAI APIPlatform
🔒 Security / Enterprise
✓ Verified enterprise information

Model available exclusively through OpenAI's cloud infrastructure (closed weights). Thinking mode and advanced features are restricted to paid plans (Plus, Pro, Business, Enterprise). API access requires OpenAI developer account verification; organizational verification may be required for full access to GPT Image models via API.

The model enforces content policy at generation time — requests that violate the policy return a 400 error (BadRequestError) with a content_policy message. AI-generated content is tagged with AI metadata. Transparent background (PNG with alpha channel) is not supported in the Responses API tool option — use gpt-image-1.5 for that purpose. Free tier access is limited to standard/instant mode with a restricted number of generations (approximately 2 images/day per tester reports). Streaming, function calling, and structured outputs are not supported by the gpt-image-2 API (confirmed on the model's page).
Updated: 26 Apr 2026↗ Security documentation

Sources and related pages

12 sources
WebIntroducing ChatGPT Images 2.0 — OpenAI official announcementDocsGPT Image 2 Model — OpenAI API DocsDocsImage Generation Guide — OpenAI API (gpt-image-2 parameters, quality, sizes)DocsOpenAI API Pricing — official token billing rates for gpt-image-2DocsOpenAI API Changelog — gpt-image-2 release entryDocsGPT Image Generation Models Prompting Guide — OpenAI Cookbook (resolution constraints, quality tiers)BlogTechCrunch — ChatGPT Images 2.0 is surprisingly good at generating textBlogThe Decoder — ChatGPT Images 2.0 thinks before it generatesBlogEngadget — ChatGPT Images 2.0 is better at rendering non-Latin textBlogWikipedia — GPT Image (rodzina modeli, architektura autoregresywna)BlogNeurohive — ChatGPT Images 2.0: Image Arena scores, API specs, benchmark breakdownBlogOfficeChai — ChatGPT Images 2.0 Tops Arena With Big Jump Over Nano Banana 2 (sub-kategorie, punktacja)
Browse related topics
📁 GPT🌐 Content generation🧠 AR Generation🧠 Native Multimodal☁ OpenAI APIAll image generation model modelsAll reasoning model models