GPT-6-Astra
OpenAICodexReasoning: low
GPT-6-Astra 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 (96.4 out of 100).
Rank 1 of 9Tied with GPT-6.1-Sol
TifinaghBench score
29.4out of 100
95% interval 28.3–30.4
28.5 with strict spelling
0–100, with 95% interval
Script score
96.4out of 100
95% interval 95.7–97.0
77.7 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
- 29.0244 correct, 93 partial, 663 wrong of 1000
- English → Tifinagh
- 11.7117 correct, 0 partial, 883 wrong of 1000
- Tifinagh ↔ Latin
- 99.0990 correct, 0 partial, 10 wrong of 1000
Tarifit
- Tifinagh → English
- 37.8333 correct, 84 partial, 576 wrong of 993
- English → Tifinagh
- 9.089 correct, 0 partial, 904 wrong of 993
- Tifinagh ↔ Latin
- 93.0930 correct, 0 partial, 70 wrong of 1000
Kabyle
- Tifinagh → English
- 62.1561 correct, 116 partial, 320 wrong of 997
- English → Tifinagh
- 26.7266 correct, 0 partial, 731 wrong of 997
- Tifinagh ↔ Latin
- 97.1971 correct, 0 partial, 29 wrong of 1000
| Tifinagh → English | English → Tifinagh | Tifinagh ↔ Latin | |
|---|---|---|---|
| Standard Moroccan Amazigh | 29.0244 correct, 93 partial, 663 wrong of 1000 | 11.7117 correct, 0 partial, 883 wrong of 1000 | 99.0990 correct, 0 partial, 10 wrong of 1000 |
| Tarifit | 37.8333 correct, 84 partial, 576 wrong of 993 | 9.089 correct, 0 partial, 904 wrong of 993 | 93.0930 correct, 0 partial, 70 wrong of 1000 |
| Kabyle | 62.1561 correct, 116 partial, 320 wrong of 997 | 26.7266 correct, 0 partial, 731 wrong of 997 | 97.1971 correct, 0 partial, 29 wrong of 1000 |
By dialect
Meaning tasks only
- Standard Moroccan Amazigh20.4
- Tarifit23.4
- Kabyle44.4
By task
- Tifinagh → English43.0
- English → Tifinagh15.8
- Tifinagh ↔ Latin96.4
Tifinagh → English grades
Free-text meanings, graded by the judge or matched exactly.
- Correct
- Partial
- Wrong
- Failed
Standard Moroccan Amazigh244 correct, 93 partial, 663 wrong of 1000
Tarifit333 correct, 84 partial, 576 wrong of 993
Kabyle561 correct, 116 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 → Tifinagh | Tifinagh ↔ Latin | |
|---|---|---|
| Standard Moroccan Amazigh | 66.3% | 0.2% |
| Tarifit | 75.0% | 17.1% |
| Kabyle | 55.3% | 0.7% |
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
- 97 answer, 30 judge
- Tokens
- 1,042,163 in · 163,647 out · 7,775 reasoning
- Model cost
- Not reported by the harness
- Judge cost
- $0.19
- Dates
- 20 Oct 2026, 10:44 UTC → 20 Oct 2026, 10:56 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 Amazigh
633505d9ccd2244f9a53b64f56bd153c428cc3dbd7229066845404bf3d7c5060 - Tarifit
d144a5619aa31cc06ffcc4eb26c361c65ca2c71451dfdfbea7fa134206481c4f - Kabyle
ed3a9882ac0f4ef10630917a302fd7d9ff6ac706654bb317b94c715f6b4cf9ea