Skip to content
TifinaghBench

ⴰⵣⵓⵍAzul! That’s hello in Amazigh.

Which AI models know Amazigh?

We test how well AI models know Amazigh words written in Tifinagh, across three dialects, with no dictionary or web search to help.

The best get about 3 in 10 right: GPT-6.1-Sol and GPT-6-Astra, with 29.9 out of 100.

Leaderboard

What is the TifinaghBench score?
How well a model knows Amazigh words, from 0 to 100. It averages two tasks in three dialects: saying what a Tifinagh word means, and writing an English word in Tifinagh. 30 means roughly 3 right answers in 10.
What does ± mean?
The margin of error. Each model answered 1,000 words per dialect, so the score could move by about this much with a different sample. Two models closer than their combined margins are tied.
What does tied mean?
The gap between these models is smaller than the margin of error, so we can’t say which one is better. They share the rank.
What is a harness?
The app the model ran in: Codex, Claude Code or OpenCode. These are the coding assistants people use every day. Tools and web search were turned off.
What is reasoning?
How long the model was allowed to think before answering. Higher settings think longer and cost more.
  1. 1tied: Shares this rank: the gap to the model it ties with is within the measurement noise.
    GPT-6.1-Sol

    OpenAI · Codex · high reasoning

    29.9± 1.0
  2. 1tied: Shares this rank: the gap to the model it ties with is within the measurement noise.
    GPT-6-Astra

    OpenAI · Codex · low reasoning

    29.4± 1.0
  3. Clear gap: 2.1 points below the group above
  4. 3
    Claude Opus 5.5

    Anthropic · Claude Code · high reasoning

    27.3± 1.0
  5. Clear gap: 9.1 points below the group above
  6. 4
    GPT-6-Luna

    OpenAI · Codex · high reasoning

    18.2± 0.8
  7. Clear gap: 1.4 points below the group above
  8. 5
    DeepSeek v4.1 Flash

    DeepSeek · OpenCode · high reasoning

    16.8± 0.8
  9. Clear gap: 3.9 points below the group above
  10. 6tied: Shares this rank: the gap to the model it ties with is within the measurement noise.
    Claude Sonnet 5.5

    Anthropic · Claude Code · high reasoning

    12.9± 0.7
  11. 6tied: Shares this rank: the gap to the model it ties with is within the measurement noise.
    GLM 5.3 Flash

    Zhipu AI · OpenCode · high reasoning

    12.0± 0.7
  12. Clear gap: 4.0 points below the group above
  13. 8
    Claude Haiku 5.5

    Anthropic · Claude Code · high reasoning

    8.0± 0.6
  14. Clear gap: 6.6 points below the group above
  15. 9
    MiniMax M3.1 Flash

    MiniMax · OpenCode · high reasoning

    1.4± 0.3

0–100, with 95% interval.