Skip to content

GPT-6.1-Sol

OpenAICodexReasoning: high

GPT-6.1-Sol 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 (96.5 out of 100).

Rank 1 of 9Tied with GPT-6-Astra

Compare with others

TifinaghBench score

29.9out of 100

95% interval 28.8–30.9

29.0 with strict spelling

0–100, with 95% interval

Script score

96.5out of 100

95% interval 95.8–97.1

77.5 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
31.0263 correct, 94 partial, 643 wrong of 1000
English → Tifinagh
11.2112 correct, 0 partial, 888 wrong of 1000
Tifinagh ↔ Latin
99.6996 correct, 0 partial, 4 wrong of 1000

Tarifit

Tifinagh → English
38.7335 correct, 98 partial, 560 wrong of 993
English → Tifinagh
8.079 correct, 0 partial, 914 wrong of 993
Tifinagh ↔ Latin
92.8928 correct, 0 partial, 72 wrong of 1000

Kabyle

Tifinagh → English
62.0560 correct, 117 partial, 320 wrong of 997
English → Tifinagh
28.4283 correct, 0 partial, 714 wrong of 997
Tifinagh ↔ Latin
97.0970 correct, 0 partial, 30 wrong of 1000

By dialect

Meaning tasks only

  • Standard Moroccan Amazigh21.1
  • Tarifit23.3
  • Kabyle45.2

By task

 

  • Tifinagh → English43.9
  • English → Tifinagh15.8
  • Tifinagh ↔ Latin96.5

Tifinagh → English grades

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

  • Correct
  • Partial
  • Wrong
  • Failed
  • Standard Moroccan Amazigh263 correct, 94 partial, 643 wrong of 1000

  • Tarifit335 correct, 98 partial, 560 wrong of 993

  • Kabyle560 correct, 117 partial, 320 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 Amazigh66.8%0.1%
Tarifit80.6%17.3%
Kabyle51.7%0.8%

The run

Setup
Official setup
Harness
Codex 0.163.0@openai/codex-sdk 0.163.0
Judge
Claude Haiku 5.5 (Claude Code)
Sample
1,000 words per dialect, seed 1
Batch size
100
Sessions
93 answer, 30 judge
Tokens
1,074,029 in · 478,011 out · 320,688 reasoning
Model cost
Not reported by the harness
Judge cost
$0.19
Dates
19 Oct 2026, 23:19 UTC → 20 Oct 2026, 00:07 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