AVGen-Bench
Models generate videos with audio from prompts across eleven real-world categories. Specialist models and multimodal judges measure perceptual quality, synchronization, and adherence to detailed visual, speech, music, and physical requirements.
The catalog supplied an unrelated paper identifier. The maker repository links arXiv 2604.08540. Synchronization metrics are lower-is-better, while the total and most component scores are higher-is-better.
- Made by
- Microsoft Research
- Released
- April 2026
- Kind
- Video generation
- Level
- University
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- Last model added
- March 2026
- Scores refreshed
- November 18th, 2026
In the authors’ words
Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture the fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and a universal breakdown in musical pitch control. Code and benchmark resources are available at http://aka.ms/avgenbench.
Seedance 2.0 tops the leaderboard with 72.1 Veo 3.1 Fast Audio leads with 67.9 among models released by March 2026
Best score per entry from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. 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 Veo 3.1 Fast Audio, at 67.9. 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:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 Seedance 2.0 72.1
- 2
Veo 3.1 Fast Audio 67.9
- 3
Veo 3.1 Audio 66.3
- 4 Sora 2 64.2
- 5
Wan 2.6 63.0
- 6 Seedance-1.5 Pro 62.5
- 7 Kling-V2.6 61.8
- 8 LTX-2.3 60.0
- 9 NanoBanana2 + MOVA 58.1
- 10 LTX-2 56.6
- 11 Emu3.5 + MOVA 56.1
- 12 Wan2.2 + HunyuanVideo-Foley 53.3
- 13 Ovi 52.0
Source: Official AVGen-Bench leaderboard
- 1
Veo 3.1 Fast Audio 67.9
- 2
Veo 3.1 Audio 66.3
- 3 Sora 2 64.2
- 4
Wan 2.6 63.0
- 5 LTX-2.3 60.0
Source: Official AVGen-Bench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Seedance 2.0 | 72.1 | Official leaderboard | ||
| 2 | | 67.9 | Official leaderboard | ||
| 3 | | 66.3 | Official leaderboard | ||
| 4 | Sora 2 | 64.2 | Official leaderboard | ||
| 5 | | 63.0 | Official leaderboard | ||
| 6 | Seedance-1.5 Pro | 62.5 | Official leaderboard | ||
| 7 | Kling-V2.6 | 61.8 | Official leaderboard | ||
| 8 | LTX-2.3 | 60.0 | Mar 2026 | Official leaderboard | |
| 9 | NanoBanana2 + MOVA | 58.1 | Official leaderboard | ||
| 10 | LTX-2 | 56.6 | Official leaderboard | ||
| 11 | Emu3.5 + MOVA | 56.1 | Official leaderboard | ||
| 12 | Wan2.2 + HunyuanVideo-Foley | 53.3 | Official leaderboard | ||
| 13 | Ovi | 52.0 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 67.9 | Official leaderboard | ||
| 2 | | 66.3 | Official leaderboard | ||
| 3 | Sora 2 | 64.2 | Official leaderboard | ||
| 4 | | 63.0 | Official leaderboard | ||
| 5 | LTX-2.3 | 60.0 | Mar 2026 | Official leaderboard |
Rankings by part
We read 9 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Text, Face and Music. Sora 2 leads 4 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.
Text
5 models we track have a score here.
- 1
Wan 2.6 ↑3 77.0
- 2
Veo 3.1 Audio 76.5
- 3
Veo 3.1 Fast Audio ↓2 75.1
- 4 Sora 2 ↓1 74.8
- 5 LTX-2.3 54.2
Face
5 models we track have a score here.
- 1
Veo 3.1 Audio ↑1 52.9
- 2
Veo 3.1 Fast Audio ↓1 52.8
- 3 Sora 2 51.2
- 4
Wan 2.6 49.3
- 5 LTX-2.3 45.1
Music
5 models we track have a score here.
- 1 Sora 2 ↑2 7.81
- 2
Veo 3.1 Audio 5.00
- 3
Veo 3.1 Fast Audio ↓2 3.13
- 4
Wan 2.6 1.75
- 5 LTX-2.3 1.38
Speech
5 models we track have a score here.
- 1
Veo 3.1 Audio ↑1 96.1
- 2
Veo 3.1 Fast Audio ↓1 94.5
- 3
Wan 2.6 ↑1 89.3
- 4 Sora 2 ↓1 88.6
- 5 LTX-2.3 86.7
Holistic
5 models we track have a score here.
- 1 Sora 2 ↑2 88.9
- 2
Veo 3.1 Fast Audio ↓1 86.3
- 3
Veo 3.1 Audio ↓1 84.1
- 4
Wan 2.6 81.0
- 5 LTX-2.3 65.2
Visual quality
Raw 0 to 1 score, multiplied by 100 in the total. 5 models we track have a score here.
- 1
Veo 3.1 Fast Audio 0.96
- 2
Wan 2.6 ↑2 0.96
- 3
Veo 3.1 Audio ↓1 0.95
- 4 LTX-2.3 ↑1 0.86
- 5 Sora 2 ↓2 0.85
Audio quality
Production quality score, multiplied by 10 in the total. 5 models we track have a score here.
- 1
Wan 2.6 ↑3 7.15
- 2 LTX-2.3 ↑3 7.11
- 3
Veo 3.1 Audio ↓1 6.77
- 4
Veo 3.1 Fast Audio ↓3 6.64
- 5 Sora 2 ↓2 5.91
Low-level kinematic plausibility
Raw 0 to 5 score, multiplied by 20 in the total. 5 models we track have a score here.
- 1 Sora 2 ↑2 4.05
- 2 LTX-2.3 ↑3 3.99
- 3
Veo 3.1 Audio ↓1 3.74
- 4
Wan 2.6 3.69
- 5
Veo 3.1 Fast Audio ↓4 3.68
High-level causal reasoning
5 models we track have a score here.
- 1 Sora 2 ↑2 79.0
- 2
Veo 3.1 Audio 68.5
- 3
Veo 3.1 Fast Audio ↓2 67.4
- 4
Wan 2.6 66.9
- 5 LTX-2.3 64.3
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official AVGen-Bench 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 Audio and video
- Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy.
Qwen 3.8 Max 88.7%
- LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization.
Gemini 4 Argon 91.7%
- MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy.
Qwen 3.8 Max 82.4%
- MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy.
Qwen 3.8 Max 90.8%
- FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better.
Phi-4-multimodal 4.00%
- VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules.
Qwen3.5-Omni-Plus 93.1%
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 8 months ago, in March 2026, for LTX-2.3. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (13 entries).
- Pass Models scored 5 models we track, plus 8 other entries on its leaderboard.
- Warning Newest score The newest came 8 months ago, in March 2026, for LTX-2.3.
- Note Frontier models None of its models are on the capabilities index.
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on microsoft.github.io: 5 rows for models we track and 8 other entries at the last read.
- Pass Last read November 18th, 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.