Skip to content

Claude Opus 5.5

AnthropicClaude CodeReasoning: high

Claude Opus 5.5 gets about 3 vocabulary questions in 10 right. It does best in Kabyle. Its weakest spot is writing Tarifit words in Tifinagh. It swaps Tifinagh and Latin letters almost perfectly (96.0 out of 100).

Rank 3 of 9

Compare with others

TifinaghBench score

27.3out of 100

95% interval 26.3–28.3

26.3 with strict spelling

0–100, with 95% interval

Script score

96.0out of 100

95% interval 95.3–96.7

77.6 with strict spelling

Transliteration score, 0–100

Scores by dialect and task

Share of the score earned in each cell, 0–100. The transliteration column measures the script, not the vocabulary.

Standard Moroccan Amazigh

Tifinagh → English
22.0175 correct, 89 partial, 736 wrong of 1000
English → Tifinagh
9.696 correct, 0 partial, 904 wrong of 1000
Tifinagh ↔ Latin
99.1991 correct, 0 partial, 9 wrong of 1000

Tarifit

Tifinagh → English
37.2314 correct, 110 partial, 569 wrong of 993
English → Tifinagh
7.070 correct, 0 partial, 923 wrong of 993
Tifinagh ↔ Latin
92.5925 correct, 0 partial, 75 wrong of 1000

Kabyle

Tifinagh → English
59.0530 correct, 117 partial, 350 wrong of 997
English → Tifinagh
29.0289 correct, 0 partial, 708 wrong of 997
Tifinagh ↔ Latin
96.5965 correct, 0 partial, 35 wrong of 1000

By dialect

Meaning tasks only

  • Standard Moroccan Amazigh15.8
  • Tarifit22.1
  • Kabyle44.0

By task

 

  • Tifinagh → English39.4
  • English → Tifinagh15.2
  • Tifinagh ↔ Latin96.0

Tifinagh → English grades

Free-text meanings, graded by the judge or matched exactly.

  • Correct
  • Partial
  • Wrong
  • Failed
  • Standard Moroccan Amazigh175 correct, 89 partial, 736 wrong of 1000

  • Tarifit314 correct, 110 partial, 569 wrong of 993

  • Kabyle530 correct, 117 partial, 350 wrong of 997

Character error rate

For exact-match answers: the smallest edit distance to an accepted answer, divided by its length. Lower is better; it shows near misses that a 0 score hides.

English → TifinaghTifinagh ↔ Latin
Standard Moroccan Amazigh70.0%0.1%
Tarifit82.1%16.9%
Kabyle51.3%0.9%

The run

Setup
Official setup
Harness
Claude Code 2.1.302@anthropic-ai/claude-agent-sdk 0.3.303
Judge
Claude Haiku 5.5 (Claude Code)
Sample
1,000 words per dialect, seed 1
Batch size
100
Sessions
100 answer, 31 judge
Tokens
735,978 in · 728,783 out · 494,495 reasoning
Model cost
$17.01 (API-equivalent; run on a subscription)
Judge cost
$0.18
Dates
20 Oct 2026, 22:36 UTC → 20 Oct 2026, 22:39 UTC
CLI version
0.1.0
Dataset checksums

SHA-256 of each dictionary file, so the run can be checked once the datasets are released.

  • Standard Moroccan Amazigh633505d9ccd2244f9a53b64f56bd153c428cc3dbd7229066845404bf3d7c5060
  • Tarifitd144a5619aa31cc06ffcc4eb26c361c65ca2c71451dfdfbea7fa134206481c4f
  • Kabyleed3a9882ac0f4ef10630917a302fd7d9ff6ac706654bb317b94c715f6b4cf9ea

Run folder on GitHub