Three Claude 5.5 models got the same end-of-service sum right: 52,750 dirhams. The test, by Mohamed Abdou, shows only how closely each model followed request details. It is three tasks, so it ranks nobody in Arabic.
Abdou runs a site selling AI skills and tools. He tested Opus 5.5, Sonnet 5.5 and Haiku 5.5 with no tools inside Claude Code. His page was published on 22 September 2026 and last updated on 15 October 2026.
The short answer
- Gratuity calculation: all three reached 52,750 dirhams. Haiku added an unrequested labor-law note. The rule was written in the prompt, so the sum tests arithmetic, not law.
- Facebook ad copy: three Egyptian-Arabic copies with a 120-character limit. All passed. Opus’s copies were 77, 90 and 97 characters, and the author called the first the least Egyptian.
- Right-to-left CSS: Opus made seven changes with reasons and raised line-height to 1.6. Haiku left it at 1.2.
- Takeaway: the cheaper model kept its numbers correct. The gap was in detail-following, not accuracy.
How was the test run?
The author says answers were graded without the grader knowing which model wrote them. He also says every count and calculation was recomputed in code.
His own conclusion is that three tasks are not enough for a general verdict. We agree. One task produced identical answers, one was a pass for all, and only the CSS task showed a visible gap.
What can’t this test tell you?
- No ranking in Arabic: it covers three tasks only.
- Dialect: Egyptian Arabic does not show quality in Saudi Arabic or formal Arabic.
- Law: the calculation does not prove a labor law or that it fits a country, because the rule was written into the prompt.
- Source: a commercial site that sells skills and tools, not a neutral measurement body.
Is there independent Arabic measurement?
Artificial Analysis’s methodology says its multilingual evaluation, which includes Arabic, is separate from its general intelligence index. We found no documented Arabic results there for these versions. We do not fill the gap with an English or overall score.
What should you do if you build for Arabic users?
- Test on a sample of your own work: your tone, dialect and text lengths.
- Recompute numbers and counts in code before relying on them.
- Do not rely on a model’s answer for a legal question without an official source.
We suggest grading a small sample of your real prompts blind, as Abdou did, before you pick a model. That is our suggestion, not a rule from the test.
Questions and answers about the Claude 5.5 Arabic test
Does it prove Opus is better than Haiku in Arabic?
No. Three tasks are not enough, and in the first one all three reached the same answer.
Does Egyptian Arabic represent all Arabic?
No. That task measures one dialect.
Does the gratuity result prove a labor law?
No. The rule was written into the prompt, so the result does not prove a law or show that it fits a country.
Where did the models differ most?
In the CSS task. Opus gave seven changes with reasons and set line-height to 1.6, while Haiku left it at 1.2.
Should I trust a model’s counts of characters?
Check them. The author recomputed every count in code, and the ad copies came in at 77, 90 and 97 characters against a 120 limit.
Terms in plain language
- End-of-service gratuity: a payment calculated for an employee when a job ends.
- Right-to-left page: a page whose text runs from right to left, as in Arabic.
- Line-height: the spacing between lines of text in CSS.
- Blind grading: grading answers without knowing which model wrote them.
Sources
- Mohamed Abdou: Opus 5.5 on three Arabic tasks (published 22 September 2026, updated 15 October 2026)
- Artificial Analysis: evaluation methodology
Facts read on 21 October 2026.
