Benchmark Almanac

VisEditBench (TikZ diagram editing)

Gemini 3.1 Pro 83.6% leads since September 2026
Not in model cards

Models edit TikZ code to follow instructions drawn on a diagram. A judge scores adherence, preservation of other content, and readability.

The authors list 29 code editing model entries for 694 diagram edits. The full-score rate counts outputs that earn five on all three image criteria.

Released
September 2026
Kind
Spatial reasoning and counting
Level
University
Scored
Graded by a model
Answers
Code
Tasks
All public
Measures
Capability
Language
English
In and out
Images, Code → Code
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
In views
Vision
Strongest model tested
GPT-5.4 #28 on the capability index
Last model added
April 2026
Scores refreshed
November 20th, 2026
25
Models scored on its own leaderboard

Gemini 3.1 Pro leads with 83.6% among models released by April 2026

Best score per model from the benchmark's own 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.

The newest model with a score here came out in April 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 Pro 83.6%
  2. 2 Gemini 3 Flash 71.8%
  3. 3 Qwen3.5 397B-A17B Thinking 66.3%
  4. 4 Claude Opus 4.6 59.9%
  5. 5 Qwen3.5-35B-A3B Thinking 53.9%
  6. 6 GPT-5.4 47.3%
  7. 7 Gemini 3.1 Flash-Lite 44.8%
  8. 8 Gemma 4 31B IT 43.5%
  9. 9 Claude Sonnet 4.5 42.4%
  10. 10 Qwen3-VL-30B-A3B-Thinking 42.1%
  11. 11 GLM-4.6V 40.5%
  12. 12 GLM-4.5V 38.3%
  13. 13 Qwen3.5-9B Thinking 38%
  14. 14 Qwen3-VL-235B-A22B-Instruct 36%
  15. 15 Qwen3-VL-235B-A22B thinking 35.7%
  16. 16 Qwen3-VL-8B-Thinking 30.8%
  17. 17 Qwen3-VL-4B-Thinking 30%
  18. 18 Qwen3.5-4B Thinking 28.1%
  19. 19 Claude Haiku 4.5 25.2%
  20. 20 Qwen3-VL-30B-A3B-Instruct 18.7%
  21. 21 GPT-5.4 Mini 18.2%
  22. 22 Qwen3-VL-8B-Instruct 17.9%
  23. 23 Qwen3-VL-4B-Instruct 12.8%
  24. 24 Gemma 4 E4B 6.05%
  25. 25 GPT-5.4 Nano 4.61%

Source: Official VisEditBench (TikZ diagram editing) leaderboard

#ModelScoreSettingDateSource
1 Gemini 3.1 Pro 83.6% Feb 2026 Official leaderboard
2 Gemini 3 Flash 71.8% Dec 2025 Official leaderboard
3 Qwen3.5 397B-A17B 66.3% Thinking Feb 2026 Official leaderboard
4 Claude Opus 4.6 59.9% Feb 2026 Official leaderboard
5 Qwen3.5-35B-A3B 53.9% Thinking Feb 2026 Official leaderboard
6 GPT-5.4 47.3% Mar 2026 Official leaderboard
7 Gemini 3.1 Flash-Lite 44.8% Mar 2026 Official leaderboard
8 Gemma 4 31B IT 43.5% Apr 2026 Official leaderboard
9 Claude Sonnet 4.5 42.4% Sep 2025 Official leaderboard
10 Qwen3-VL-30B-A3B-Thinking 42.1% Oct 2025 Official leaderboard
11 GLM-4.6V 40.5% Dec 2025 Official leaderboard
12 GLM-4.5V 38.3% Aug 2025 Official leaderboard
13 Qwen3.5-9B 38% Thinking Feb 2026 Official leaderboard
14 Qwen3-VL-235B-A22B-Instruct 36% Sep 2025 Official leaderboard
15 Qwen3-VL-235B-A22B 35.7% thinking Sep 2025 Official leaderboard
16 Qwen3-VL-8B-Thinking 30.8% Oct 2025 Official leaderboard
17 Qwen3-VL-4B-Thinking 30% Oct 2025 Official leaderboard
18 Qwen3.5-4B 28.1% Thinking Feb 2026 Official leaderboard
19 Claude Haiku 4.5 25.2% Oct 2025 Official leaderboard
20 Qwen3-VL-30B-A3B-Instruct 18.7% Oct 2025 Official leaderboard
21 GPT-5.4 Mini 18.2% Mar 2026 Official leaderboard
22 Qwen3-VL-8B-Instruct 17.9% Oct 2025 Official leaderboard
23 Qwen3-VL-4B-Instruct 12.8% Oct 2025 Official leaderboard
24 Gemma 4 E4B 6.05% Mar 2026 Official leaderboard
25 GPT-5.4 Nano 4.61% Mar 2026 Official leaderboard

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Qwen3.5 397B-A17B
Thinking
66.3%
Qwen3.5-35B-A3B
Thinking
53.9%
Qwen3.5-9B
Thinking
38%
Qwen3.5-4B
Thinking
28.1%

Source: Official VisEditBench (TikZ diagram editing) leaderboard

Number one over time

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

Google DeepMind
Own leaderboard Not tested since Sep 2026

Source: Official VisEditBench (TikZ diagram editing) leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Gemini 3.1 Pro Gemini 3 Flash Claude Sonnet 4.5 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Gemini 3.1 Pro Gemini 3 Flash

Source: Official VisEditBench (TikZ diagram editing) 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

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 7 months ago, in April 2026, for Gemma 4 31B IT. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (25 entries).
  • Pass Models scored 25 models we track.
  • Warning Newest score The newest came 7 months ago, in April 2026, for Gemma 4 31B IT.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.4 (#28).
  • Pass Headroom The best is 83.6%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on viseditbench.github.io: 29 rows for models we track 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