AI Image Models in 2026: The Year in Review

Key takeaways

• OpenAI ends the year on top. GPT Image 2.5, released September 8, holds first place on both the LMArena and Artificial Analysis text-to-image leaderboards, with Microsoft, Google, and xAI close behind.

• Text rendering became the headline feature. Release after release led on typography: readable product names, prices, translated copy, and dense layouts that used to need a designer to fix.

• Editing by reference replaced most masking. Leading models now take up to 10 reference images and change one element of a picture while leaving the rest untouched.

• 4K became a standard API option, and cheap images got cheaper. A 1K image from Google's newest model fell from $0.134 on Nano Banana Pro to $0.0336 on Nano Banana 2.1, while OpenAI's top-quality images got more expensive.

• Open weights got closer but not level. The best open model sits 19th on LMArena and 21st on Artificial Analysis, and two of the strongest open releases are licensed for non-commercial use only.

• The rules arrived. New York's synthetic performer law took effect on June 9, EU watermarking duties start this year, and Meta, Amazon, and Walmart all tightened their AI image rules.

The year at a glance

Six things that changed

1. Text rendering became the headline feature

At the start of the year, a model that spelled a product name right on the first try was a pleasant surprise. By the fall, it's the first claim in most launch posts. Google led Nano Banana 2 with "precision text rendering and translation." ByteDance showed Seedream 5.0 Pro rendering a dense sale poster without spelling errors. xAI said Grok Imagine Image 2.0 plans typography and layout "the way a designer would," and Microsoft reported its biggest gain for MAI-Image-2.6 on text. LMArena now scores text rendering as a separate category.

For ad teams, that moves a whole class of work into the model: price callouts, multilingual versions of a static ad, packaging mockups, and infographic-style product images. Those claims come from the makers, so proof every word before an image goes out, but the error rate is no longer the blocker it was.

2. Reference editing replaced most masking

The biggest workflow change wasn't generation at all; it was how images get edited. FLUX 3 Image, FLUX 3 Image Klein, and Qwen-Image-2.1 each take up to 10 reference images; Grok Imagine Image 2.0 takes five; and Nano Banana 2 keeps up to five characters and 14 objects consistent across a scene. Instead of painting a mask, you describe the change or draw a box: FLUX 3 Image accepts bounding boxes in the prompt and makes pixel-exact local edits, Seedream 5.0 Pro edits by point, lasso, or sketch, and GPT Image 2.5 changes a single element without disturbing the subject. Midjourney folded its separate reference tools into one Edit Model in July.

Masks haven't gone away, and some models still take them. For product work, though, the practical result is that one approved packshot can be dropped into many scenes with a far better chance that the label survives.

3. 4K arrived in the APIs

GPT Image 2 outputs up to 3840 pixels, Nano Banana 2 and 2.1 go to 4K, FLUX 3 Image reaches about 16 megapixels with custom sizes up to an 8:1 ratio, and Reve 2.1 renders 4K x 4K natively. Some well-known models are still 2K-class: Midjourney's HD mode is 2048 pixels, as are Ideogram 4 and Recraft V4 Pro.

4K costs more. On FLUX 3 Image, a 4K image is $0.607 against $0.048 at 1K, and on Nano Banana 2.1 it's $0.113 against $0.0336. For marketplace listings that top out around 2,000 pixels, 2K is usually enough.

4. Cheap images got cheaper, and the best got pricier

At the low end, prices fell fast. Google's 1K image went from $0.134 on Nano Banana Pro (released in November 2025) to $0.067 on Nano Banana 2 in February and $0.0336 on Nano Banana 2.1 in October, and 2.1 scores higher than 2 on both major leaderboards. Recraft's cheapest model fell from $0.04 to $0.007 per image, and the open FLUX 3 Image Klein runs $0.009 per 1K image through Black Forest Labs' own API.

OpenAI went the other way at the top. A high-quality 1024 x 1024 image cost $0.133 on GPT Image 1.5 and costs $0.211 on GPT Image 2, while the low-quality tier fell from $0.009 to $0.006. On Artificial Analysis, the top 21 models now range from $9 to $211 per 1,000 images. The gap makes it worth testing a cheaper model for high-volume variants and keeping the expensive one for hero images.

5. Open weights got closer, with license strings attached

Open-weight image models had their best year. Black Forest Labs released FLUX.2 [klein] in January and FLUX 3 Image Klein in October, Ideogram published its first open model, Alibaba shipped Qwen-Image-2.1, and HiDream released HiDream-O1-Image under MIT.

They still trail the closed models. On the Artificial Analysis leaderboard, the best open model, FLUX 3 Image Klein, is 21st; the top 20 are all proprietary. On LMArena's text-to-image board, Qwen-Image-2.1 is the best open model at 19th, with Ideogram 4 20th.

The licenses matter more than the rankings for ad work. Qwen-Image-2.1 is for "research or evaluation purposes only," a change from earlier Qwen-Image releases under Apache 2.0, and Ideogram 4 is non-commercial. FLUX 3 Image Klein is free for commercial use by individuals and companies under $5 million in annual revenue. Only some open models can go into a paid campaign without a separate license.

6. Models started researching before they draw

A newer trend is generation grounded in live information. Nano Banana 2 can ground images in web search results, Seedream 5.0 Lite has optional real-time web search, Microsoft added web grounding to MAI-Image-2.6, and Meta's Muse Image uses search and code tools and checks its own output. For ads, that helps with current products, landmarks, and real-world references, and it also raises the stakes on checking what the model pulled in.

Where the leaderboards ended up

LMArena's text-to-image leaderboard on October 26, 2026, based on 6.87 million votes:

Artificial Analysis has the same top three, followed by Nano Banana 2.1, MAI-Image-2.6, and Grok Imagine Image 2.0. On LMArena's image editing board, OpenAI's two 2.5 models also lead, with Microsoft's MAI-Image-2.7 preview third on a small number of votes.

The other story is Microsoft. MAI-Image-2 shipped in March, five more versions followed, and by October MAI-Image-2.6 was fourth on LMArena, ahead of Google and xAI on that board.

The rules caught up

New York. The synthetic performer disclosure law (GBL 396-b) took effect on June 9, 2026. Advertisers who know an ad uses a synthetic performer, a digitally created person meant to pass as a human performer without being any real, identifiable one, have to disclose it conspicuously, with penalties of $1,000 for a first violation and $5,000 for each after.

The EU. Article 50 of the AI Act, which requires providers of generative AI to mark its output in a machine-readable way, applies from August 2, 2026. A July amendment, Regulation (EU) 2026/1744, gives generative systems already on the market before that date until December 2, 2026, to comply.

Meta. Since a June 1, 2026, update, Meta labels ads with "AI info" in "About this ad" and says it will detect ads made or edited with third-party AI tools through industry-standard signals.

Marketplaces. Amazon added a contains-synthetic-performer tag for listing images and video that show photorealistic AI-generated people, and Walmart requires AI images to be "truthful, accurate and not misleading." Our guides to Amazon and Walmart image requirements cover both.

Provenance. OpenAI's images carry C2PA metadata and an invisible watermark, Google pairs its SynthID watermark with C2PA Content Credentials, and in iOS 27.2 Apple added SynthID to most images made or edited with Apple Intelligence. Assume any AI image you publish can be identified as one.

What was retired

OpenAI removed DALL-E 2 and DALL-E 3 from its API on May 12 and shut down the original GPT Image 1 on October 23. GPT Image 1.5, GPT Image 1 Mini, and the ChatGPT image endpoint follow on December 1, according to OpenAI's deprecations page. Google shut down Imagen 4 on August 17, with Nano Banana 2.1 as the replacement, and xAI retired its Grok Imagine Pro image model in May. If an automated ad workflow still calls one of these, it has already broken or will soon.

What it means for ad teams

• Retest your text-heavy formats. Price cards, multilingual statics, and packaging mockups that failed a year ago may now work in one pass.

• Rebuild product workflows around references. One approved packshot plus a reference-based model beats re-generating the product every time.

• Price by tier. Use a cheap, fast model for volume variants and a top model for hero images; the price gap is now 20 times or more.

• Check the license on open models. Several of the best are non-commercial or carry revenue limits.

• Plan for disclosure. Between New York, the EU, Meta, and the marketplaces, AI people and AI-made product images need a disclosure plan before launch, not after.

Frequently asked questions

What is the best AI image model in 2026?

On both LMArena and Artificial Analysis at the end of October, OpenAI's GPT Image 2.5 Sunburst ranks first, followed by GPT Image 2.5 Flare and GPT Image 2. Microsoft's MAI-Image-2.6, Google's Nano Banana 2.1, and xAI's Grok Imagine Image 2.0 round out the top six.

What is the best open-weight image model in 2026?

It depends on the leaderboard. Artificial Analysis ranks FLUX 3 Image Klein highest among open models (21st overall), and LMArena ranks Qwen-Image-2.1 highest (19th overall). Qwen-Image-2.1 is licensed for research and evaluation only, so FLUX 3 Image Klein is the stronger choice for commercial work, free under $5 million in annual revenue.

Which AI image models can render text accurately?

Text rendering was a headline feature for Nano Banana 2 and 2.1, Seedream 5.0 Pro, Grok Imagine Image 2.0, MAI-Image-2.6, and Ideogram 4, and GPT Image 2.5 leads the overall leaderboards. These are the makers' claims, so proofread every image before it runs.

Is DALL-E still available?

No. OpenAI removed DALL-E 2 and DALL-E 3 from its API on May 12, 2026. Its current image models are GPT Image 2 and GPT Image 2.5.

Do AI-generated images in ads need to be labeled?

In some places, yes. New York requires conspicuous disclosure of synthetic performers in ads from June 9, 2026, the EU AI Act requires machine-readable marking of AI-generated content, Meta adds "AI info" labels to ads, and Amazon has a tag for photorealistic AI people in listings.

How much does an AI-generated image cost in 2026?

Through the makers' own APIs, from under a cent to about 60 cents. Google's Nano Banana 2.1 is $0.0336 for a 1K image, FLUX 3 Image Klein is $0.009, and a high-quality 1024 x 1024 image from GPT Image 2 is $0.211. A 4K image from FLUX 3 Image costs $0.607.