Every Arabic-language list of "AI sites for programming" we could find was written in 2024, still recommends tools that no longer exist under those names, and never asks the one question an Arabic-speaking developer actually has: if I type my request in Arabic, will the tool understand it, and will I be able to read what comes back? We spent 25 October 2026 in the documentation, public bug trackers and published benchmarks of eleven coding assistants to answer it. The short version: the models understand Arabic well; the editors and terminals they live in mostly cannot display it.
Arabic prompts work, at a measurable cost
Three studies put numbers on the question. HumanEval-XL, published at LREC-COLING in May 2024, translated the standard code-generation problems into 23 natural languages including Arabic; the leading OpenAI model of the time scored 82.5 on Python with English prompts and 73.75 with Arabic ones, while a weaker model fell from 66.25 to 52.50. The 3LM benchmark from Abu Dhabi's Technology Innovation Institute, presented at the Arabic NLP conference in November 2025, produced human-checked Arabic versions of HumanEval and MBPP and found a much smaller gap for current open models, about four to five points, with Arabic and English scores correlated at 0.97: a model that codes well from English codes well from Arabic. A January 2026 paper by Saudi researchers, the first Arabic benchmark for tool calling, found agent accuracy drops five to ten percent when the user speaks Arabic, which is directly relevant to the agentic coding tools below.
Vendors are more guarded than researchers. GitHub's documentation calls English "the primary supported language" for Copilot Chat, warns of "differing performance with prompts in other languages", and states that commit messages and pull-request summaries are generated in English only. Google's Gemini Code Assist is the single product whose documentation names Arabic as a supported prompt language, one of 38, on a page last updated in June 2026. Anthropic documents a language setting for Claude Code that forces replies into any language you name, Arabic included, and keeps working after long sessions are compacted, although permission dialogs stay in English. OpenAI, Cognition, Replit, Lovable, Bolt and Vercel publish nothing about prompt language at all.
The real failure is rendering, and it is nobody's bug
The problem you will hit first is not wrong code. It is Arabic text coming back reversed, with letters separated, or with the paragraph aligned to the wrong side. The cause sits below every tool: xterm.js, the terminal engine inside VS Code and therefore inside Cursor, Antigravity and the Copilot and Claude Code panels, has an open request for right-to-left support that dates from 2017 and is still labelled "help wanted". VS Code's own editor has had an open issue on right-to-left text since 2016. Every command-line agent inherits this.
The public trackers show the pattern. OpenAI's Codex CLI has a bug open since July 2026 for Arabic rendered as separated, reordered characters on Windows, with no maintainer reply. Anthropic's Claude Code repository holds more than a dozen Arabic and right-to-left reports filed between November 2025 and September 2026, most closed as "not planned"; a fresh one opened on 18 October 2026 describes reversed Arabic in a PhpStorm terminal on Windows and notes that Codex displays the same text correctly on the same machine. Google's Gemini CLI closed a universal bidirectional-text request in May 2026 along with the pull request that would have implemented it, saying there were no immediate plans. Cursor staff described garbled Arabic in the chat panel as "a known, long-standing issue" in February and again in June 2026. OpenAI support, asked in June and July 2026 for right-to-left support in the ChatGPT and Codex apps, gave the same answer twice: no promise, no timeline.
What this means in practice: letter shaping is usually right, but word order, mixed Arabic-English lines and paragraph alignment are not. Arabic comments inside a source file display in the editor but scramble when they share a line with Latin identifiers. The working habits that appear again and again in those bug reports are sensible ones: prompt in Arabic in the chat panel if you prefer, keep identifiers and function names in English, put user-facing Arabic strings in resource files rather than inline, and review output in a browser rather than a terminal. Community extensions exist that patch the direction in VS Code, Cursor and Antigravity for Claude Code, Copilot and Gemini Code Assist; none is official and all can break on the next update.
Tool by tool: what is documented, and what it costs
Cursor. Acquired by SpaceX in a deal that closed in August 2026 and repriced in October 2026 into tiers bundled with Grok and X: Free, Plus at $8, Premium at $30, Super at $100 and Ultra at $200 a month, differing mainly in usage allowance. Documentation is available in twelve languages, none of them Arabic; the editor UI can be switched through VS Code language packs, but the agent panel stays in English.
GitHub Copilot. A free tier with 2,000 completions a month, then Pro at $10, Pro+ at $39 and Max at $100 a month, metered in AI credits. GitHub Docs are translated into nine languages, not Arabic. Custom instructions can force Arabic replies; commit-message generation cannot be changed from English.
Claude Code. Included with Pro at $20 a month and Max from $100. Saudi Arabia is on Anthropic's supported-countries list. The language setting is documented; the docs themselves come in eleven languages without Arabic, and this is the tool with the longest public record of Arabic rendering reports, none resolved at the root.
OpenAI Codex and ChatGPT. Free includes limited Codex access; Go is $8 in the United States with pricing that "varies by market"; Plus is $20; Pro starts at $100. Saudi Arabia is on OpenAI's supported list and the company runs an Arabic-language website, but the Codex documentation says nothing about prompt language or text direction.
Devin Desktop. This is Windsurf's name since 2 June 2026, after the Cognition acquisition; the old Cascade agent was retired in July. Free, Pro at $20, Max at $200, Teams at $80 plus $40 per developer. We found no mention of Arabic anywhere in its documentation.
Google. The lineup changed completely in 2026: Gemini CLI became Antigravity CLI in May, individual access to Gemini CLI and Gemini Code Assist ended on 18 June, and Code Assist sales to businesses stopped on 9 October with service continuing into 2027 for existing subscribers. Antigravity is free for individuals with weekly limits, raised by Google AI Pro and Ultra subscriptions. The irony is that the only document naming Arabic belongs to the product individuals can no longer reach, and Antigravity's documentation has no equivalent list. Jules, the asynchronous agent, offers 15 tasks a day free and says nothing about language.
Replit. Core at $20 and Pro at $100 a month. In January 2026 Arabic.AI Academy launched free Arabic-language Replit courses in partnership with the company, the first full localisation of its teaching material, but this is education, not a localised product; the documentation remains English.
Lovable, Bolt and v0. The prompt-to-app builders: Lovable from $25 a month, Bolt at $18 a month on annual billing, v0 at $30. None documents Arabic support. Lovable's January 2026 guide to multilingual websites does warn against ignoring right-to-left languages, but that concerns the apps you build, not the builder.
The Saudi and Gulf picture
There is no Arabic-native coding assistant yet. HUMAIN launched HUMAIN Chat on the ALLaM model in August 2025 and the HUMAIN ONE agent platform in 2026 without mentioning code in either, and the small ALLaM model scored about 31 on Arabic code in 3LM against roughly 80 for the best open models. The movement is in Abu Dhabi: the Jais family's release of 15 October 2026 ships Apache-licensed open weights and, for the first time for an Arabic-centred model, agentic coding numbers, 70.2 on SWE-Bench Verified and 73.8 on Terminal-Bench, with an agent platform announced as "coming soon". Qatar's Fanar publishes no code results. Searching in Arabic for a Saudi or Gulf startup offering a coding assistant returned nothing.
Paying from Saudi Arabia
Every price above is in US dollars as the vendors publish it. When we opened the subscription pages from an account signed in from Saudi Arabia on 25 October 2026, two of them switched to riyals: ChatGPT Plus at SAR 90 a month and Google AI Pro at SAR 85.99, while Claude stayed in dollars. Saudi Arabia appears explicitly on the supported-country lists of OpenAI, Anthropic and Google. Which local payment methods each one accepts is covered in our guide to paying for AI tools from Saudi Arabia.
What we tested ourselves: Claude Code on Linux
On 25 October 2026 our team ran one hands-on session with Claude Code in the one environment for which we had found no documented report at all: Ubuntu with the default GNOME Terminal, an English-locale system, and a Max subscription. The tool was at version 2.1.309 when we started and updated itself to 2.1.310 between the two rounds.
Comprehension passed every step. An Arabic prompt asking for a Python function that averages a list while ignoring empty values, with Arabic comments in the code and a three-sentence explanation, produced a correct function on the first attempt (we ran it), Arabic comments as requested, an Arabic explanation, and, unprompted, an Arabic one-line summary from the tool itself. We then set the documented language option to Arabic and wrote the next prompts in English: every reply came back in Arabic, and it stayed Arabic after we compacted the conversation, which some write-ups say pushes the tool back to English. After compaction it wrote nine unit tests, ran them and explained the result in Arabic. We did not test permission dialogs because the session ran in auto mode.
Rendering was better than the Windows and macOS reports and worse than acceptable on mixed lines. Letters join correctly and word order is right on every purely Arabic line; the Arabic comments inside the code read without trouble because each contains a single Latin token at its end. A line that mixes Arabic with four or more Latin tokens becomes a puzzle: each Arabic stretch is internally correct, but the stretches are laid out left to right, so a bold heading lands mid-line, list numbers sit at the far left, and a closing None drops to the far right where an Arabic reader begins. The paragraph is being treated as English; the letters are not being garbled. We believe the intact shaping owes to the VTE engine behind GNOME Terminal, which implements bidirectional text where xterm.js does not, but that is an inference, not a measurement.
Not tested: Windows, macOS, Cursor, Codex, Copilot, and any terminal other than GNOME's. It is one result on one machine on one day, but it is the first screenshot-backed account of Arabic in Claude Code on Linux we know of.
What we could not verify
We found no dated Arabic-language review that actually tests Arabic prompts in Cursor, Copilot, Claude Code or Codex; every Arabic page we read was a translation or a feature comparison. We could not extract per-language Arabic figures from the mHumanEval appendix, find ALLaM's HumanEval score in its own paper, locate any statement on Arabic from Devin Desktop, Replit, Lovable, Bolt or v0, or find an official riyal price for any tool. Our own testing was limited to Claude Code on Linux, as described above.
What to choose
If Arabic is your prompting language, pick a tool that documents it: Gemini Code Assist if your company holds a licence, Claude Code with the language setting, or Copilot with custom instructions. If you write a lot of Arabic comments and strings, work in web interfaces or the browser rather than a terminal, and expect to patch text direction with a community extension. If you want something open and built in the Gulf, watch for the Jais agent platform. Whatever you choose, keep the code in English and the prompts in whatever language you think in; that is what the numbers support.