GPT Image 2 fits teams that need reliable multilingual text in generated images and are inside OpenAI's stack, replacing gpt-image-1.5 and DALL-E 3 across ChatGPT and the API. It ranks #1 on the LM Arena image leaderboard with a 9.6 out of 10 rating, but the reasoning pipeline pushes high-quality latency past 50 seconds, ruling out live or interactive use.
Released April 21, 2026, GPT Image 2 is OpenAI's O-series reasoning image model, the first commercial image API to plan and review each generation across four stages before producing pixels. It takes up to 16 reference images per request and natively outputs 2K resolution, extending to 4096x4096 pixels for print-ready work.
Where it sits
- $13.50/M$ per 1M tokensBlended price (3:1)Lower is better#59 / 64peer median $1.70/Mvendor price, checked by HokAI
- --tokens/sOutput speedHigher is better-- / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- --% correctGPQA DiamondHigher is better-- / 44peer median 88.3%per source, see benchmark scores
Pricier than 92% of the 64 GA models with a published price, and one of 22 that document a zero-data-retention option.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: OpenAI · Family: GPT Image
Input modalities: text, image · Output: image
About GPT Image 2
GPT Image 2 (API model ID gpt-image-2) is OpenAI's third generation image model, branded ChatGPT Images 2.0 in the consumer product and released April 21, 2026. It replaces gpt-image-1 (April 2025) and gpt-image-1.5 (December 2025), both deprecated June 2, 2026 with API removal set for December 1, 2026; DALL-E 2 and DALL-E 3 were removed from the API on May 12, 2026. The model's break from its predecessors is an O-series reasoning pipeline: a four-stage Understand, Plan, Generate, Review sequence runs before any pixel is produced, rather than a single-pass denoising loop. OpenAI has not disclosed whether the generation stage itself is diffusion, autoregressive, or a Transfusion-style hybrid, only that it is a new architecture; the model remains fully proprietary with no released weights.
Each request takes a text prompt, reference images, and an optional mask, and returns raster images only across five aspect ratios from square to widescreen, with a maximum output of 4096x4096 pixels; there is no context window in the LLM sense and no function calling or structured output. Editing sessions are stateful: follow-up natural language instructions refine a prior generation without re-uploading the base image, alongside pixel-precise mask-based control on the same editing endpoint.
GPT Image 2 holds the top rank on the LM Arena image generation leaderboard as of June 2026, rated 9.6 out of 10 by independent reviewers across prompt adherence, text accuracy, editing, and commercial suitability; text rendering accuracy improved sharply over gpt-image-1's 60-70% baseline, particularly for CJK and Arabic scripts.
The API is available directly through OpenAI and via Microsoft Azure AI Foundry, which layers Azure AI Content Safety on top of OpenAI's own moderation; as of June 2026 it is not offered on AWS Bedrock or Google Vertex AI. OpenAI published a dedicated system card for ChatGPT Images 2.0 on release day, covering pre- and post-generation content classifiers and additional deepfake safeguards given the model's higher photorealism.
Generated images carry full commercial rights, with API inputs retained for a limited window for abuse monitoring by default and enterprise zero-retention agreements available on request. The model is SOC 2 Type 2 compliant through OpenAI's enterprise tier and HIPAA-eligible for qualifying customers, with a training cutoff of December 2025 supplemented at inference by the built-in web search. OpenAI has not published a roadmap for the next gpt-image release; prior releases have shipped roughly every six to nine months.
Pricing
$8 per 1M image input tokens, $30 per 1M image output tokens, $2 per 1M cached image input tokens, and $5 per 1M text input tokens. At 1024x1024, per-image cost is about $0.006 at low quality, $0.053 at medium, and $0.211 at high. The Batch API cuts every rate by 50% for asynchronous delivery within 24 hours. There is no free tier.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.240 | $0.030 | $0.270 |
| Support reply | $0.016 | $0.0090 | $0.025 |
| One coding agent run | $1.60 | $0.600 | $2.20 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- O-Series Reasoning Pipeline: Runs a four-stage Understand, Plan, Generate, Review sequence before committing to pixels, replacing the single-pass denoising loop most diffusion image models use.
- 99% Text Rendering Accuracy: Renders text at roughly 99% accuracy across Latin, CJK, and Arabic scripts in the same request, the highest reported figure among commercial image APIs as of June 2026.
- 16 Reference Images Per Request: Accepts up to 16 reference images alongside the prompt for consistent characters, products, and brand styling across a single generation, four times the limit on gpt-image-1.
- Natural Language Inpainting and Outpainting: Edits or extends an image from a plain-language instruction with no manual mask required for most jobs, though a precise mask stays available on the same editing endpoint.
- Built-In Web Search: Queries the live web before generating to supplement the December 2025 training cutoff, useful for time-sensitive or real-world reference requests without a manual workaround.
Pros
- Ranked #1 on the LM Arena image leaderboard as of June 2026, the clearest independent quality signal available for a model this new.
- Multilingual text rendering is a genuine step up from every prior gpt-image release, particularly for CJK and Arabic scripts that most competitors still garble.
- Natural language inpainting plus multi-image reference support cuts the prompt-engineering overhead of mask-based editors for consistent product and character shots.
Cons
- High-quality output can take up to 50 seconds because of the four-stage reasoning pass, ruling out real-time or interactive workflows.
- Token-based billing makes per-image cost hard to predict in advance; teams should model spend with OpenAI's calculator rather than a flat per-image estimate.
- Closed weights rule out self-hosting, and the model is not available on AWS Bedrock or Google Vertex AI as of June 2026.
Benchmarks
- LMArena rank: #1 vendor-reported · 24 Jun 2026 — Position on the blind human-preference leaderboard; #1 is best.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does GPT Image 2 actually cost?
Pricing is token-based rather than flat per-image: OpenAI charges $8 per 1M image input tokens, $2 per 1M for cached input, $5 per 1M text input tokens, and $30 per 1M image output tokens. Real-world cost lands around $0.006 for a low-quality 1024x1024 image, $0.053 at medium quality, and $0.211 at high quality, and the Batch API cuts every one of those rates in half if you can wait up to 24 hours for results. GPT Image 2 has no free tier, so testing requires a paid OpenAI or Azure account.
Does GPT Image 2 have a free plan?
GPT Image 2 skips a free tier entirely: every request, from the first call, runs through a paid OpenAI or Azure API account. Teams testing the model typically start at the low-quality setting, the cheapest of the three tiers, before scaling up to medium or high quality for production assets.
What are the best alternatives to GPT Image 2?
FLUX.2 from Black Forest Labs leads on photorealism and exact brand-color matching rather than text accuracy. Gemini 3 Pro Image (Nano Banana Pro) adds Google Search-grounded facts to each generation but costs more per image at medium-to-high settings. Riverflow 2.0 Pro topped Artificial Analysis in both generation and editing and bills a flat per-image rate instead of tokens, which suits teams that want predictable costs over OpenAI's reasoning-heavy pipeline.
What separates GPT Image 2 from Riverflow 2.0 Pro?
Riverflow 2.0 Pro topped Artificial Analysis in text-to-image and editing ahead of OpenAI's prior flagship and FLUX.2 [max], and it bills a flat rate per image rather than by token, which makes budgeting simpler. GPT Image 2 answers with its own O-series reasoning pipeline and the higher multilingual text accuracy of the two, plus direct distribution through the OpenAI API and Azure AI Foundry that many teams already hold credentials for. Neither vendor has published a head-to-head benchmark between the two models specifically, so teams with strict text-rendering requirements should still test both against their own prompts.
How do you get started with GPT Image 2?
Create an OpenAI API key, or an Azure AI Foundry deployment, and call the generation endpoint with a text prompt and, optionally, several reference images. Start at quality='low' to keep costs down while testing prompts, since the four-stage reasoning pass only adds meaningful latency at medium and high quality. The official Python and Node.js SDKs support gpt-image-2 without extra libraries, and the Batch API is worth switching to once a workflow produces more than a handful of images per run.
Top Alternatives
- FLUX.2: Pick FLUX.2 if you need exact hex-color brand matching and leading photorealism; pick GPT Image 2 if the image must contain accurate multilingual text.
- Gemini 3 Pro Image: Pick Gemini 3 Pro Image if you need Google Search-grounded facts baked into the picture; pick GPT Image 2 for materially cheaper medium-quality output.
- Riverflow 2.0 Pro: Pick Riverflow 2.0 Pro if flat per-image pricing and its Artificial Analysis lead over OpenAI's prior flagship matter most; pick GPT Image 2 for direct OpenAI and Azure API access.