Skip to content

Claude Haiku 5.5

AnthropicClaude CodeReasoning: high

Claude Haiku 5.5 gets about 1 vocabulary question in 10 right. It does best in Kabyle. Its weakest spot is writing Standard Moroccan Amazigh words in Tifinagh. It still makes mistakes swapping Tifinagh and Latin letters (81.3 out of 100).

Rank 8 of 9

Compare with others

TifinaghBench score

8.0out of 100

95% interval 7.4–8.6

7.9 with strict spelling

0–100, with 95% interval

Script score

81.3out of 100

95% interval 80.0–82.6

70.1 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
3.923 correct, 32 partial, 945 wrong of 1000
English → Tifinagh
0.88 correct, 0 partial, 992 wrong of 1000
Tifinagh ↔ Latin
89.0890 correct, 0 partial, 110 wrong of 1000

Tarifit

Tifinagh → English
8.562 correct, 44 partial, 887 wrong of 993
English → Tifinagh
0.99 correct, 0 partial, 984 wrong of 993
Tifinagh ↔ Latin
61.9619 correct, 0 partial, 381 wrong of 1000

Kabyle

Tifinagh → English
25.4202 correct, 102 partial, 693 wrong of 997
English → Tifinagh
8.484 correct, 0 partial, 913 wrong of 997
Tifinagh ↔ Latin
93.0930 correct, 0 partial, 70 wrong of 1000

By dialect

Meaning tasks only

  • Standard Moroccan Amazigh2.4
  • Tarifit4.7
  • Kabyle16.9

By task

 

  • Tifinagh → English12.6
  • English → Tifinagh3.4
  • Tifinagh ↔ Latin81.3

Tifinagh → English grades

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

  • Correct
  • Partial
  • Wrong
  • Failed
  • Standard Moroccan Amazigh23 correct, 32 partial, 945 wrong of 1000

  • Tarifit62 correct, 44 partial, 887 wrong of 993

  • Kabyle202 correct, 102 partial, 693 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 Amazigh88.4%3.2%
Tarifit94.5%19.9%
Kabyle77.1%1.8%

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
93 answer, 30 judge
Tokens
784,886 in · 1,170,057 out · 923,891 reasoning
Model cost
$0.69 (API-equivalent; run on a subscription)
Judge cost
$0.22
Dates
20 Oct 2026, 14:05 UTC → 20 Oct 2026, 14:30 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