TutorBench
Tutoring tasks for high-school and AP subjects. Scale AI.
As of 25 Oct 2026, Muse Spark leads TutorBench on BenchLeader with 68.5%, ahead of GPT-5.4 Pro at 56.6%, across 27 model configurations with a published result.
- Published by
- Scale AI SEAL
- Category
- Instruction following
- Index weight
- 0.5
- Models
- 27
- Data as of
- 25 Oct 2026
- Newest result
- 8 Apr 2026
The publisher has posted no new results on this board since 8 Apr 2026. BenchLeader re-reads it every day, so newer models appear here as soon as they are published.
Scale AI SEAL Leaderboards.
What the test looks like
TutorBench asks the model to act as a tutor for high-school and AP subjects: explaining, giving hints, and grading student work, judged against rubrics.
How it is scored
Percent of rubric points achieved, published by Scale AI.
What to keep in mind
Rubric-graded pedagogy, which is subjective; it rewards restraint (hinting) over simply giving the answer.
- 1Muse Spark68.5%
- 2GPT-5.4 Pro56.6%
- 3Gemini 2.5 Pro55.6%
- 4GPT-555.3%
- 5o3-pro54.6%
- 6Kimi K2.554.6%
- 7GPT-5.1 (thinking)54.1%
- 8Claude Opus 4.6 (max)53.7%
- 9Gemini 3 Pro53.7%
- 10Claude Opus 4.6 (none)53.5%
- 11GPT-5.253.5%
- 12Gemini 3.1 Pro53.0%
- 13o3 (medium)52.8%
- 14o3 (high)52.1%
- 15Gemini 3.1 Flash Lite51.5%
| # | ||||
|---|---|---|---|---|
| 1 |  Muse Spark | 68.5% | 64.3 | 2026-04-08 |
| 2 |  GPT-5.4 Pro | 56.6% | 63.3 | 2026-03-23 |
| 3 |  Gemini 2.5 Pro | 55.6% | 53.9 | 2025-09-10 |
| 4 | GPT-5 | 55.3% | 56.2 | 2025-09-10 |
| 5 | o3-pro | 54.6% | 53.4 | 2025-09-10 |
| 6 |  Kimi K2.5open ↗ | 54.6% | 54.5 | 2026-02-12 |
| 7 | GPT-5.1thinking | 54.1% | 54.1 | 2025-11-26 |
| 8 |  Claude Opus 4.6max | 53.7% | 57.4 | 2026-02-17 |
| 9 | Gemini 3 Pro | 53.7% | 60.4 | 2025-09-10 |
| 10 | Claude Opus 4.6no reasoning | 53.5% | 56.5 | 2025-11-21 |
| 11 | GPT-5.2 | 53.5% | 58.0 | 2025-12-15 |
| 12 | Gemini 3.1 Pro | 53.0% | 60.5 | 2026-03-25 |
| 13 | o3medium | 52.8% | 53.3 | 2025-09-10 |
| 14 | o3high | 52.1% | 53.9 | 2025-11-26 |
| 15 | Gemini 3.1 Flash Lite | 51.5% | 53.7 | 2026-03-23 |
| 16 | Claude Opus 4.5thinking | 51.2% | 59.9 | 2025-09-10 |
| 17 | Claude Opus 4.1thinking | 50.8% | 56.4 | 2025-11-26 |
| 18 | Claude Opus 4.5 | 49.8% | 57.9 | 2025-09-10 |
| 19 | Claude Opus 4thinking | 49.7% | 55.0 | 2025-10-08 |
| 20 | GPT 5.1 Instant | 49.1% | – | 2025-09-10 |
| 21 | Claude Sonnet 4.5thinking | 49.0% | 52.6 | 2025-09-10 |
| 22 | Claude Opus 4.1 | 47.4% | 51.5 | 2025-11-26 |
| 23 | Claude 37 Sonnetthinking | 46.5% | – | 2025-10-08 |
| 24 | Claude Sonnet 4.5 | 45.7% | 53.7 | 2025-09-10 |
| 25 | Claude Opus 4 | 45.5% | 52.6 | 2025-09-10 |
| 26 | Llama 4 Maverickopen ↗ | 40.2% | 42.8 | 2025-09-10 |
| 27 | GPT-4o | 36.1% | 43.9 | 2026-02-17 |
More instruction following benchmarks · All LLM benchmarks
Cite as: BenchLeader, “TutorBench leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/scale_tutorbench.
TutorBench: questions
- What does TutorBench measure?
- TutorBench asks the model to act as a tutor for high-school and AP subjects: explaining, giving hints, and grading student work, judged against rubrics. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads TutorBench?
- Muse Spark leads TutorBench with 68.5% as of 25 Oct 2026, ahead of GPT-5.4 Pro at 56.6%.
- How many models have TutorBench results?
- 27 model configurations have a TutorBench result on BenchLeader, all taken from Scale AI SEAL.
- Who runs TutorBench and how often is it updated?
- TutorBench is published by Scale AI SEAL. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does TutorBench count toward the BenchLeader Index?
- Yes. TutorBench contributes to the instruction following category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.