benchr
Type "a coffee-shop menu board" into an image model a year ago and you got a beautiful sign covered in melted gibberish. Letters that almost spelled words. That single failure kept image generators out of real design work, where the text is the whole point. GPT Image 2, released April 21, 2026, is the first model where that complaint mostly goes away.
OpenAI ships it two ways. Inside ChatGPT it's branded "Images 2.0" and runs on every plan, including the free tier. In the API it's the model gpt-image-2, and it replaces the now-deprecated gpt-image-1. Same model underneath, two front doors. This review reads the public evidence: OpenAI's image-generation docs, the pricing page, the Images 2.0 system card, and the public preference leaderboards.
The text problem, mostly solved
The headline capability is rendering legible text, and not just a three-word logo. OpenAI's own framing points at "generating detail and complexity such as dense text," and that's where you feel the jump: a slide with a title and four bullets, an infographic with labels, a packaging mockup with a real ingredient list, signage in a non-Latin script. Earlier models fell apart as soon as the word count climbed. This one holds together through dense layouts and multilingual text.
Two more changes matter in daily use. There's an optional "thinking" mode for image generation on paid plans, which spends more compute to plan a harder composition before it draws. And generation is up to around four times faster than the previous model, so the iterate-and-tweak loop that image work depends on is less painful.
Where it still slips
Don't read "solved" as "perfect." OpenAI's guide is candid that the model "can still struggle with precise text placement and clarity," especially on busy prompts, and that it can be inconsistent holding visual elements in a structured layout. In practice you'll still get the occasional dropped letter or a caption nudged half a line off. The fix is the same as with any generator: treat the first output as a draft, read every word, and regenerate the ones it muffs.
There are hard limits worth knowing before you wire it into a pipeline. GPT Image 2 doesn't do transparent backgrounds, which rules out a lot of logo and sticker work without a separate cutout step. And on the API, streaming, function calling, structured outputs, and fine-tuning aren't supported for the image model, so it doesn't slot into agent loops the way a text model does.
Does it win the leaderboards?
On Arena.ai's blind text-to-image preference test, where people pick the better of two unlabeled images, GPT Image 2 sits at #1 across the board with a wide lead over the next model. Take the exact Elo with a grain of salt; the live figure read inconsistently when we checked, and crowd-preference scores drift as votes accumulate. The reliable claim is the ranking and the size of the gap, not a precise number. This is the same caution benchr's piece on why benchmarks stopped telling you anything raises about every leaderboard: a #1 badge is a starting point, not a verdict on your specific prompts.
Worth being clear on scope, too. This review is about image generation, the model drawing a picture. That's a different question from how well a model reads and reasons about an image you give it, which is what benchr's multimodal capability ranking measures across the frontier models. A model can be great at one and ordinary at the other.
What it costs
Inside ChatGPT, there's no separate charge: image generation is bundled into your plan, free tier included, with usage caps OpenAI doesn't publish as a fixed number. That makes it the easy default for one-off images.
For volume, the API is where the math matters. A 1024x1024 image runs roughly $0.006 at low quality, about $0.053 at medium, and around $0.211 at high, with token costs layered on top (about $5 per million text-input tokens, $8 per million image-input tokens, and $30 per million image-output tokens). Those per-image figures shift by size and quality, so price your real settings against OpenAI's calculator before you budget a batch. API access also needs organization verification, a step that trips up first-time users. If you're already weighing the broader OpenAI stack, the GPT-5.5 review covers where the text side of that lineup landed.
The verdict
Go with ChatGPT Images 2.0 if words appear anywhere in your image. Nothing else on the market handles dense, correct text as reliably right now, and that one capability covers most of the practical design work people want from a generator. The free access inside ChatGPT makes it worth trying before you consider a paid image tool.
Hold off, or keep a second tool around, if you live in transparent-background assets, need tight programmatic control through the API, or work in a style where text never shows up and another model's look suits you better. For everyone making slides, signage, and social graphics, the gibberish era is over, and this is the model that ended it.
Frequently asked
What is ChatGPT Images 2.0 and how is it different from GPT Image 2?
They're the same release. "ChatGPT Images 2.0" is how OpenAI markets image generation inside ChatGPT; "GPT Image 2" (model ID gpt-image-2) is the underlying model in the API. It launched April 21, 2026 and replaces the older gpt-image-1, which is deprecated.
Can GPT Image 2 render readable text in images?
Yes, and that's its headline improvement. It handles dense text layouts like infographics, slides, menus, and diagrams, plus multilingual scripts, far better than earlier image models. OpenAI still cautions it can struggle with precise text placement and clarity on complex prompts, so proof the output before you ship it.
What does ChatGPT Images 2.0 cost?
In ChatGPT it's included in your plan, including the Free tier, with usage limits OpenAI doesn't publish as a fixed number. Via the API (gpt-image-2), each image runs roughly $0.006 (low) to $0.211 (high) at 1024x1024, on top of token costs of about $5 per million text-input tokens and $30 per million image-output tokens. API access requires organization verification.
What are GPT Image 2's main weaknesses?
Per OpenAI's own guide, text placement and clarity can still slip, complex prompts add latency, and it can be inconsistent holding visual elements in structured layouts. It also doesn't support transparent backgrounds, and streaming, function calling, structured outputs, and fine-tuning aren't available for the image model.
Is GPT Image 2 the best image generator right now?
It sits at #1 on Arena.ai's blind text-to-image preference leaderboard by a wide margin, though the exact Elo reads inconsistently and these are crowd-sourced votes, not a fixed score. For text-heavy work it's the clear pick. For other styles, run your own prompts against a rival before committing.
Changelog
- May 30, 2026 — Originally published. Capabilities, pricing, limits, and leaderboard standing verified against OpenAI's image-generation docs, pricing page, the Images 2.0 system card, and Arena.ai.
References
- OpenAI, "gpt-image-2 model," developers.openai.com, accessed May 2026.
- OpenAI, "Image generation guide," developers.openai.com, accessed May 2026.
- OpenAI, "Pricing," developers.openai.com, accessed May 2026.
- OpenAI, "ChatGPT Images 2.0 system card," deploymentsafety.openai.com, April 21, 2026.
- Arena.ai, "Text-to-Image leaderboard," arena.ai, May 2026 snapshot.