Ask almost any AI image generator for a shop sign that reads "مرحباً" and you will usually get something that looks Arabic from across the room and falls apart up close: letters that should join stand alone, the word runs left to right, a dot migrates and changes the meaning. English text in generated images became a solved problem between 2023 and 2026. Arabic did not come along for the ride. We read the official documentation and pricing pages of thirteen image generators, two multilingual benchmarks published in 2026, and the dated tests we could find in Arabic-language blogs and vendor forums. The short version: three vendors put Arabic in writing, three steer you away from it in their own help pages, and the rest say nothing at all.
The mechanics of the failure
Arabic is cursive by design. Most letters change shape depending on where they sit in a word, the line runs right to left, and dots and diacritics distinguish letters that are otherwise identical. An image model does not apply those rules; it reproduces patterns from its training images, which are dominated by English and Chinese. Alibaba's technical report on its Qwen-Image model says so directly: most systems are trained predominantly on English or Chinese glyphs and struggle with accurate characters, consistent spacing or the correct reading order in other scripts. Nihad Nadam, the Dubai-based calligrapher behind ArabicDesign.ai, puts it another way: a generator learns visual relationships, so it can produce a texture that resembles Arabic calligraphy without knowing whether the word is still correct.
Two 2026 studies measured the gap. MultiTextEdit, a benchmark of image-editing models across twelve languages, found that Hebrew and Arabic show the largest declines of any script, with right-to-left outputs often keeping the layout while scrambling the character order. LingT2I, published in August 2026 across seventeen models, calls non-Latin writing systems a persistent bottleneck. Neither publishes a per-model Arabic ranking, and no public leaderboard does.
Three vendors that name Arabic
Google is the only large Western vendor whose documentation lists Arabic. Its Firebase documentation for the Gemini image models, updated 14 October 2026, names Arabic among the prompt languages that the newest generation handles best and states that text inside images is supported in the same languages. The same page says the older Nano Banana model supports in-image text in English only, which means the answer to "can Gemini write Arabic?" depends on which tier you are using inside the app. Google's Middle East blog announced legible text "in multiple languages including Arabic" for Nano Banana Pro in November 2025.
ByteDance makes the most specific claim we found. The Seedream model blog of 8 July 2026 says the model adapts to the typographic rules of each language and can correctly process the right-to-left cursive script of Arabic, among more than ten languages, while conceding room to improve in finer-grained text. We found no independent Arabic test of it.
OpenAI is a mixed case. The April 2026 launch page for the current generation of ChatGPT Images lists Arabic in the caption of a multi-script poster and talks about improved multilingual support. Its December 2025 release notes had listed multilingual text as a known limitation. On OpenAI's own developer forum, a user reported on 26 April 2026 that the model now writes Arabic correctly but arranges the page with left-to-right logic; OpenAI's support account replied on 8 September 2026 that there was no timeline to share.
Three vendors that steer you away
Ideogram built its reputation on text in images and advertises 95 percent text accuracy against 30 to 50 percent for competitors. Its documentation is nonetheless candid: text comes out most accurately in English, non-Latin scripts such as Arabic or Chinese render less reliably, and the advice is to keep non-Latin text to a word or two. The June 2026 release was marketed as rendering text across languages, but the only benchmark in the accompanying technical post is an English OCR score. An Arabic blog that re-tested Ideogram in April 2026 on billboard prompts reported distorted or unintelligible letters. Our own practice matches that reading: the hero images on dalili.tools news articles are generated with Ideogram under a prompt that forbids any text in the image, and the Arabic headline is added afterwards as a separate layer.
Midjourney's text-generation documentation says text works best with the standard Latin alphabet and names no other script. Adobe Firefly makes no in-image text claim for its generator at all, and the Generative Text Edit feature in Firefly Boards, documented in March 2026, requires Latin characters and states that non-Latin character sets are not supported.
The silent majority
Black Forest Labs promises high-accuracy text "in multiple languages" for its FLUX models without naming one. Stability AI's latest Stable Diffusion release lists improved typography, and its model card notes that testing was conducted primarily in English. Recraft, Leonardo.Ai, Magnific (the name Freepik adopted in April 2026) and Canva's built-in generator publish no language list for rendered text. Silence is not evidence of failure, but it means nobody is promising you anything.
What the dated tests actually show
The independent evidence is thin and mostly comes from Arabic-language blogs testing a phrase or two. One test on the-aicity found that the word "الذكاء" lost letters in several attempts on the lighter Gemini image tier and came out complete on the Pro tier. A Ramadan design guide dated 19 February 2026 recorded "رمضان كريم" emerging with scattered or reversed letters. An Arabic comparison from 16 April 2026 ranked Google's models first but wrote that they usually fail on long sentences and cannot be relied on for commercial copy. On OpenAI's forum, users in March 2026 reported missing letters, wrong joins and misplaced diacritics. The pattern is the same everywhere: one or two words may land, a full sentence still needs checking letter by letter, and the same tool varies from one run to the next.
What works today
- Generate without text, then add the Arabic as a layer. This is the consensus of the Arabic guides and the vendors' own best-practice pages. Canva ships Arabic fonts and a "mirror page" option for right-to-left layouts; Adobe Express lists Arabic among its supported languages as of September 2026; Photoshop needs its World-Ready Layout composer switched on before Arabic letters will join.
- Edit rather than regenerate. Ideogram lets you select a region and repaint a wrong word, and can convert rendered text into editable layers. ChatGPT and Gemini support follow-up edits, so one bad word can be fixed without rolling the whole image again.
- Write the prompt for the script. Put the exact text in quotation marks, keep it short, and say explicitly that it must run right to left with connected letters. Google's documentation recommends generating the text first, then asking for an image that carries it.
- For calligraphy, use a tool built for it. ArabicDesign.ai, from Dubai, generates Naskh, Thuluth, Diwani, Ruq'ah and Kufic compositions and exports PNG and SVG. An independent review in July 2026 noted minor joining flaws in long sentences and a diacritic error in roughly one output in four.
Pricing, October 2026
Gemini's Saudi page lists a free plan that includes image creation, Google AI Plus at SAR 18.99 a month and AI Pro at SAR 85.99. ChatGPT offers limited image creation free and Plus at $20. Ideogram gives eligible free accounts a variable weekly credit allowance; Plus is $20 a month. Midjourney has no free tier; Basic is $10. Adobe Firefly has a free plan with limited daily generations and Standard at $9.99. Leonardo.Ai gives 150 fast tokens a day free, with Essential at $12. Canva's free plan includes up to 20 AI uses; Pro is $144 a year. Magnific's Premium plan is $20 a month, with no free tier shown on its pricing page. FLUX is sold per image with no subscription, and its open-weight models are free for non-commercial use. Stable Diffusion weights are free under a community licence for organisations under $1 million in revenue. Seedream and Qwen-Image are reached through cloud and third-party platforms, and neither has a single documented price page.
What we could not verify
We found no dated Arabic test with published images for Midjourney, FLUX, Recraft, Leonardo, Magnific, Firefly or Seedream, no per-model Arabic figures in either benchmark, no readable prices on Recraft's pricing page, and no statement on whether Ideogram's text layers support right-to-left type. We did not run generation tests of our own for this piece. If you have an image with correct Arabic text from any of these tools, with the prompt and the date, it is better evidence than anything we read.