Benchmark Almanac

Synthetic Sheet Music Reasoning Benchmark

gemini-3.1-flash 95.5% tops its leaderboard Claude Opus 4.6 90.5% leads the models we track
Not in model cards

Models answer questions about rhythm, chords, intervals, and scales. Each test question appears in text notation and as an image.

The authors provide 1,600 test questions per format. The Legato 2 team ran Gemini 3.1 Pro and GPT-5 on the visual test.

Released
September 2025
Kind
Documents, charts and text in images
Level
School
Scored
Checked automatically
Answers
Multiple choice
Tasks
All public
Measures
Capability
Subject
Arts
Language
English
In and out
Text, Images → Text
Interaction
Single turn
Items from
Generated
Refresh
One fixed set
Run by
Anyone
In views
Vision
Strongest model tested
GPT-5.4 #28 on the capability index
Last model added
March 2026
Scores refreshed
November 20th, 2026
13
Entries on MuTheoryEval’s leaderboard, 8 models we track

gemini-3.1-flash tops the leaderboard with 95.5% Claude Opus 4.6 leads with 90.5% among models released by March 2026

Best score per entry from MuTheoryEval’s leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Claude Opus 4.6, at 90.5%. This view shows only models we track, each at its best.

The newest model with a score here came out in March 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 gemini-3.1-flash lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ 95.5%
  2. 2 Claude Opus 4.6 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 90.5%
  3. 3 gemini-3.1-flash (minimal) lite, temperature=0, max_output_tokens=4096, thinking={"level":"minimal"}, answer_format=MCQ 89%
  4. 4 Claude Sonnet 4.6 lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ 84%
  5. 5 deepseek-reasoner lite, max_output_tokens=4096, thinking=native, answer_format=MCQ 80%
  6. 6 Gemini 3.1 Flash-Lite lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ 78%
  7. 7 GPT-5.4 lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 67.5%
  8. 8 Qwen3-Max lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 67%
  9. 9 GLM-5 lite, max_output_tokens=4096, thinking=enabled, answer_format=MCQ 66%
  10. 10 Claude Haiku 4.5 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 63.5%
  11. 11 deepseek-chat lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 59%
  12. 12 GPT-5.4 Mini lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 54%
  13. 13 glm-z1-flash lite, max_output_tokens=4096, thinking=native, answer_format=MCQ 50.5%

Source: MuTheoryEval’s leaderboard

#EntryScoreSettingReleasedRun onSource
1 gemini-3.1-flash 95.5% lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ Apr 2026 Official leaderboard
2 Claude Opus 4.6 90.5% lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Feb 2026 Apr 2026 Official leaderboard
3 gemini-3.1-flash (minimal) 89% lite, temperature=0, max_output_tokens=4096, thinking={"level":"minimal"}, answer_format=MCQ Apr 2026 Official leaderboard
4 Claude Sonnet 4.6 84% lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ Feb 2026 Apr 2026 Official leaderboard
5 deepseek-reasoner 80% lite, max_output_tokens=4096, thinking=native, answer_format=MCQ Apr 2026 Official leaderboard
6 Gemini 3.1 Flash-Lite 78% lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ Mar 2026 Apr 2026 Official leaderboard
7 GPT-5.4 67.5% lite, max_output_tokens=1024, thinking=native, answer_format=MCQ Mar 2026 Apr 2026 Official leaderboard
8 Qwen3-Max 67% lite, max_output_tokens=1024, thinking=native, answer_format=MCQ Sep 2025 Apr 2026 Official leaderboard
9 GLM-5 66% lite, max_output_tokens=4096, thinking=enabled, answer_format=MCQ Feb 2026 Apr 2026 Official leaderboard
10 Claude Haiku 4.5 63.5% lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Oct 2025 Apr 2026 Official leaderboard
11 deepseek-chat 59% lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Apr 2026 Official leaderboard
12 GPT-5.4 Mini 54% lite, max_output_tokens=1024, thinking=native, answer_format=MCQ Mar 2026 Apr 2026 Official leaderboard
13 glm-z1-flash 50.5% lite, max_output_tokens=4096, thinking=native, answer_format=MCQ Apr 2026 Official leaderboard
#ModelScoreSettingReleasedRun onSource
1 Claude Opus 4.6 90.5% lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Feb 2026 Apr 2026 Official leaderboard
2 Claude Sonnet 4.6 84% lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ Feb 2026 Apr 2026 Official leaderboard
3 Gemini 3.1 Flash-Lite 78% lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ Mar 2026 Apr 2026 Official leaderboard
4 GPT-5.4 67.5% lite, max_output_tokens=1024, thinking=native, answer_format=MCQ Mar 2026 Apr 2026 Official leaderboard
5 Qwen3-Max 67% lite, max_output_tokens=1024, thinking=native, answer_format=MCQ Sep 2025 Apr 2026 Official leaderboard
6 GLM-5 66% lite, max_output_tokens=4096, thinking=enabled, answer_format=MCQ Feb 2026 Apr 2026 Official leaderboard
7 Claude Haiku 4.5 63.5% lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Oct 2025 Apr 2026 Official leaderboard
8 GPT-5.4 Mini 54% lite, max_output_tokens=1024, thinking=native, answer_format=MCQ Mar 2026 Apr 2026 Official leaderboard

Rankings by part

We read 9 parts from MuTheoryEval’s leaderboard, each ranking models on one part of the test, such as BarLinePlacementQuestion, ChordIdentificationQuestion and ChordKeyRootIdentificationQuestion. Claude Opus 4.6 leads 5 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

BarLinePlacementQuestion

8 models we track have a score here.

  1. 1 Claude Opus 4.6 100%
  2. 2 Gemini 3.1 Flash-Lite ↑1 87%
  3. 3 GPT-5.4 ↑1 82.6%
  4. 4 Claude Sonnet 4.6 ↓2 73.9%
  5. 5 Qwen3-Max 69.6%
  6. 6 GLM-5 69.6%
  7. 7 GPT-5.4 Mini ↑1 34.8%
  8. 8 Claude Haiku 4.5 ↓1 26.1%

The best score, 90.5%, costs $0.17 for the full run

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Anthropic Google DeepMind Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 Cost of the full run, log scale Claude Opus 4.6, lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Claude Sonnet 4.6, lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ Gemini 3.1 Flash-Lite, lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ 0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 Cost of the full run, log scale Claude Opus 4.6, lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Claude Sonnet 4.6, lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ

Source: MuTheoryEval’s leaderboard

ModelEffortHarnessScoreCost of the full runOutput tokens
Claude Opus 4.6 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 90.5% $0.17 4
Claude Sonnet 4.6 lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ 84% $0.10 189
Gemini 3.1 Flash-Lite lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ 78% $0.008 1
Claude Sonnet 4.6 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 71.5% $0.10 4
GPT-5.4 lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 67.5% $0.08 4
Qwen3-Max lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 67% $0.03 1
GLM-5 lite, max_output_tokens=4096, thinking=enabled, answer_format=MCQ 66% $0.03 1,963
Claude Haiku 4.5 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 63.5% $0.03 4
GLM-5 lite, temperature=0, max_output_tokens=1024, thinking=false, answer_format=MCQ 63.5% $0.03 2
GPT-5.4 Mini lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 54% $0.03 6

The best score, 90.5%, costs $0.17 for the full run

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Anthropic Google DeepMind Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 Cost of the full run, log scale Claude Opus 4.6, lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Claude Sonnet 4.6, lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ Gemini 3.1 Flash-Lite, lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ 0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 Cost of the full run, log scale Claude Opus 4.6, lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ Claude Sonnet 4.6, lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ

Source: MuTheoryEval’s leaderboard

ModelEffortHarnessScoreCost of the full runOutput tokens
Claude Opus 4.6 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 90.5% $0.17 4
Claude Sonnet 4.6 lite, temperature=0, max_output_tokens=8256, thinking={"budget_tokens":8000}, answer_format=MCQ 84% $0.10 189
Gemini 3.1 Flash-Lite lite, temperature=0, max_output_tokens=4096, thinking={"level":"default"}, answer_format=MCQ 78% $0.008 1
Claude Sonnet 4.6 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 71.5% $0.10 4
GPT-5.4 lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 67.5% $0.08 4
Qwen3-Max lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 67% $0.03 1
GLM-5 lite, max_output_tokens=4096, thinking=enabled, answer_format=MCQ 66% $0.03 1,963
Claude Haiku 4.5 lite, temperature=0, max_output_tokens=16, thinking=false, answer_format=MCQ 63.5% $0.03 4
GLM-5 lite, temperature=0, max_output_tokens=1024, thinking=false, answer_format=MCQ 63.5% $0.03 2
GPT-5.4 Mini lite, max_output_tokens=1024, thinking=native, answer_format=MCQ 54% $0.03 6

2 models have held number one on MuTheoryEval’s leaderboard since its release in September 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Qwen Anthropic
MuTheoryEval leaderboard Passed 85% in Feb 2026, by Claude Opus 4.6

Source: MuTheoryEval’s leaderboard

Claude Opus 4.6 on the official leaderboard passed 85% in February 2026

That was 5 months after its release. By the same age, 14% of the benchmarks from 2025 had passed 85%. All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On MuTheoryEval’s leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Opus 4.6 Qwen3-Max 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Opus 4.6 Qwen3-Max

Source: MuTheoryEval’s leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Vision

All 636 in Vision
  • MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. Qwen 3.6 Plus 86%
  • MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
  • CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
  • MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy. Seed2.1 Pro 90.7%
  • MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy. Qwen 3.8 Max 97.7%
  • MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%

Data audit

All 8 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
  • Pass Independent runs Scores from the official leaderboard (13 entries).
  • Pass Models scored 8 models we track, plus 5 other entries on its leaderboard.
  • Note Newest score The newest came 8 months ago, in March 2026, for GPT-5.4 Mini. That’s expected for a saturated test.
  • Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is GPT-5.4 (#28).
  • Note Headroom Saturated. The best is 90.5%, past 85% since February 2026.
Leaderboard import 3 pass
  • Pass Importer We read the board on MuTheoryEval: 10 rows for models we track and 5 other entries at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check