Turkish Last Exam - Sub-answer Accuracy: leaderboard
Turkish Last Exam (TLE, Töz AI, 2026): 41 native Turkish expert tasks (22 text-only, 19 with an image) built from Türkiye-specific primary sources such as the Resmî Gazete, mevzuat.gov.tr and federation rulebooks; each task is a chain of 2 to 34 graded sub-answers (497 in total) that runs from reading to recalling an official Turkish rule, computing and deciding. Unlike Humanity's Last Exam's English academic questions, TLE "contains no translated questions", and each task turns on "a rule that exists in an official Turkish source but differs from web lore, foreign practice or common sense". Protocol: "One attempt per task. No tools, no web access, no system prompt." The model ends with a "Cevap: <answers>" line that grader.py compares with answer keys computed by code, with no LLM judge; the card does not state reasoning settings, and each score is a single pilot run. Score: sub-answer accuracy (%), the share of the 497 graded sub-answers that are correct, whether or not the rest of the task's chain is. Higher is better. Source: https://huggingface.co/datasets/TozAI/Turkish-Last-Exam
Metric: Sub-answer accuracy (%). Source: huggingface.co. Saturation forecast: Around July 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6.1 Sol | 69.01 | #7 |
| 2 | GPT-6 Astra | 68.21 | #6 |
| 3 | Claude Opus 5.5 | 60.56 | #3 |
| 4 | Gemini 3.7 Flash | 57.55 | #24 |
| 5 | Claude Sonnet 5.5 | 44.47 | #14 |
No result here: #9 Gemini 4 Argon, #10 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=turkish-last-exam-sub-answer-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.