GeoDrive-Bench
Models choose driving actions from road images and local traffic conventions. The benchmark covers six countries and omits explicit country labels.
- Made by
- Yingzi Ma and colleagues
- Released
- June 2026
- Kind
- Embodied reasoning
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use
- 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
GPT-5.4 leads with 91.9% among models released by March 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 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 GPT-5.4 Rule-Given 91.9%
- 2
Qwen3-VL-235B-A22B Rule-Given 90.8%
- 3
InternVL3-8B Rule-Given 81.3%
- 4
Qwen3-VL-8B-Instruct Rule-Given 80.2%
- 5
Gemma 3 12B Rule-Given 76%
- 6
Qwen2.5-VL-7B Rule-Given 73.1%
- 7
Llama 3.2 11B Rule-Given 68.9%
- 8 llava-v1.6-mistral-7b Rule-Given 62.8%
- 9
InternVL3.5-8B Rule-Given 53.8%
Source: Official GeoDrive-Bench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.4 | 91.9% | Rule-Given | Mar 2026 | Official leaderboard |
| 2 | | 90.8% | Rule-Given | Sep 2025 | Official leaderboard |
| 3 | | 81.3% | Rule-Given | Apr 2025 | Official leaderboard |
| 4 | | 80.2% | Rule-Given | Oct 2025 | Official leaderboard |
| 5 | | 76% | Rule-Given | Mar 2025 | Official leaderboard |
| 6 | | 73.1% | Rule-Given | Jan 2025 | Official leaderboard |
| 7 | | 68.9% | Rule-Given | Sep 2024 | Official leaderboard |
| 8 | llava-v1.6-mistral-7b | 62.8% | Rule-Given | Jan 2024 | Official leaderboard |
| 9 | | 53.8% | Rule-Given | Aug 2025 | Official leaderboard |
Rankings by part
We read 24 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Perception: CN, Perception: US and Perception: UK. GPT-5.4 leads 13 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Perception: CN
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 93.2%
- 2 GPT-5.4 ↓1 91.1%
- 3
InternVL3-8B 75.8%
- 4
Qwen2.5-VL-7B ↑2 73.2%
- 5
Qwen3-VL-8B-Instruct ↓1 65.8%
- 6
Gemma 3 12B ↓1 61.6%
- 7
InternVL3.5-8B ↑2 51.6%
- 8
Llama 3.2 11B ↓1 46.3%
- 9 llava-v1.6-mistral-7b ↓1 27.4%
Perception: US
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 99%
- 2 GPT-5.4 ↓1 97.5%
- 3
Qwen2.5-VL-7B ↑3 96.4%
- 4
InternVL3-8B ↓1 95.9%
- 5
Qwen3-VL-8B-Instruct ↓1 95.9%
- 6
Gemma 3 12B ↓1 83.8%
- 7
Llama 3.2 11B 81.7%
- 8 llava-v1.6-mistral-7b 65.5%
- 9
InternVL3.5-8B 47.2%
Perception: UK
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 95.3%
- 2 GPT-5.4 ↓1 94.5%
- 3
Qwen3-VL-8B-Instruct ↑1 86.2%
- 4
Qwen2.5-VL-7B ↑2 85.8%
- 5
InternVL3-8B ↓2 84.2%
- 6
Gemma 3 12B ↓1 68%
- 7
Llama 3.2 11B 50.6%
- 8
InternVL3.5-8B ↑1 46.2%
- 9 llava-v1.6-mistral-7b ↓1 28.9%
Perception: JP
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 96.4%
- 2 GPT-5.4 ↓1 95.6%
- 3
InternVL3-8B 94.2%
- 4
Gemma 3 12B ↑1 92.3%
- 5
Qwen3-VL-8B-Instruct ↓1 88.3%
- 6 llava-v1.6-mistral-7b ↑2 85.3%
- 7
Llama 3.2 11B 79.4%
- 8
Qwen2.5-VL-7B ↓2 71.4%
- 9
InternVL3.5-8B 20.6%
Perception: SG
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 92.5%
- 2
InternVL3-8B ↑1 82.3%
- 3 GPT-5.4 ↓2 81.7%
- 4
Qwen2.5-VL-7B ↑2 78%
- 5
Qwen3-VL-8B-Instruct ↓1 75.8%
- 6
Llama 3.2 11B ↑1 71%
- 7
Gemma 3 12B ↓2 63.4%
- 8
InternVL3.5-8B ↑1 54.3%
- 9 llava-v1.6-mistral-7b ↓1 46.8%
Perception: IND
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 99.5%
- 2 GPT-5.4 ↓1 99.5%
- 3
Qwen2.5-VL-7B ↑3 97.6%
- 4
Qwen3-VL-8B-Instruct 92.2%
- 5
InternVL3-8B ↓2 91.2%
- 6
Gemma 3 12B ↓1 89.3%
- 7
Llama 3.2 11B 79.5%
- 8 llava-v1.6-mistral-7b 63.9%
- 9
InternVL3.5-8B 50.2%
Prediction: CN
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 91.4%
- 2 GPT-5.4 ↓1 88.2%
- 3
Qwen3-VL-8B-Instruct ↑1 80.6%
- 4
InternVL3-8B ↓1 75.8%
- 5 llava-v1.6-mistral-7b ↑3 72.6%
- 6
Llama 3.2 11B ↑1 72.6%
- 7
Gemma 3 12B ↓2 72.6%
- 8
Qwen2.5-VL-7B ↓2 72%
- 9
InternVL3.5-8B 59.1%
Prediction: US
9 models we track have a score here.
- 1 GPT-5.4 84.9%
- 2
InternVL3-8B ↑1 82.8%
- 3
Qwen3-VL-235B-A22B ↓1 81.2%
- 4
Qwen3-VL-8B-Instruct 81.2%
- 5
Qwen2.5-VL-7B ↑1 75.8%
- 6 llava-v1.6-mistral-7b ↑2 74.2%
- 7
Gemma 3 12B ↓2 73.7%
- 8
Llama 3.2 11B ↓1 72%
- 9
InternVL3.5-8B 51.1%
Prediction: UK
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 82.6%
- 2 GPT-5.4 ↓1 80.7%
- 3
Qwen3-VL-8B-Instruct ↑1 70.8%
- 4 llava-v1.6-mistral-7b ↑4 68.9%
- 5
InternVL3-8B ↓2 67.1%
- 6
Qwen2.5-VL-7B 61.5%
- 7
Llama 3.2 11B 59%
- 8
InternVL3.5-8B ↑1 59%
- 9
Gemma 3 12B ↓4 55.9%
Prediction: JP
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 91%
- 2 GPT-5.4 ↓1 90.5%
- 3
InternVL3-8B 86.2%
- 4
Qwen3-VL-8B-Instruct 86.2%
- 5 llava-v1.6-mistral-7b ↑3 82.9%
- 6
Qwen2.5-VL-7B 82.4%
- 7
Llama 3.2 11B 81.4%
- 8
Gemma 3 12B ↓3 81%
- 9
InternVL3.5-8B 74.3%
Prediction: SG
9 models we track have a score here.
- 1 GPT-5.4 84.6%
- 2
Qwen3-VL-235B-A22B 83.4%
- 3
Qwen3-VL-8B-Instruct ↑1 82.2%
- 4
InternVL3-8B ↓1 80.5%
- 5
Qwen2.5-VL-7B ↑1 75.1%
- 6
Gemma 3 12B ↓1 75.1%
- 7 llava-v1.6-mistral-7b ↑1 69.8%
- 8
Llama 3.2 11B ↓1 66.3%
- 9
InternVL3.5-8B 45.6%
Prediction: IND
9 models we track have a score here.
- 1 GPT-5.4 92%
- 2
Qwen3-VL-235B-A22B 88.6%
- 3
InternVL3-8B 79.5%
- 4
Qwen3-VL-8B-Instruct 77.8%
- 5
Gemma 3 12B 76.7%
- 6
Llama 3.2 11B ↑1 70.5%
- 7 llava-v1.6-mistral-7b ↑1 64.8%
- 8
Qwen2.5-VL-7B ↓2 64.8%
- 9
InternVL3.5-8B 40.9%
Planning: CN
9 models we track have a score here.
- 1 GPT-5.4 90.1%
- 2
Qwen3-VL-235B-A22B 86.8%
- 3
Gemma 3 12B ↑2 86.5%
- 4
Qwen3-VL-8B-Instruct 79.5%
- 5
InternVL3-8B ↓2 79.2%
- 6
Qwen2.5-VL-7B 78.2%
- 7
Llama 3.2 11B 76.9%
- 8
InternVL3.5-8B ↑1 75.6%
- 9 llava-v1.6-mistral-7b ↓1 70%
Planning: US
9 models we track have a score here.
- 1 GPT-5.4 91.5%
- 2
Qwen3-VL-235B-A22B 87.2%
- 3
Gemma 3 12B ↑2 84%
- 4
InternVL3-8B ↓1 83%
- 5
Qwen2.5-VL-7B ↑1 80.9%
- 6
Qwen3-VL-8B-Instruct ↓2 79.3%
- 7
Llama 3.2 11B 77.1%
- 8 llava-v1.6-mistral-7b 72.3%
- 9
InternVL3.5-8B 60.6%
Planning: UK
9 models we track have a score here.
- 1 GPT-5.4 79.7%
- 2
Qwen3-VL-235B-A22B 78.1%
- 3
Gemma 3 12B ↑2 66.8%
- 4
InternVL3-8B ↓1 65.8%
- 5
InternVL3.5-8B ↑4 65.2%
- 6
Qwen2.5-VL-7B 62%
- 7
Qwen3-VL-8B-Instruct ↓3 59.9%
- 8 llava-v1.6-mistral-7b 56.1%
- 9
Llama 3.2 11B ↓2 55.6%
Planning: JP
9 models we track have a score here.
- 1 GPT-5.4 89.8%
- 2
Qwen3-VL-235B-A22B 88.2%
- 3
Gemma 3 12B ↑2 85%
- 4
InternVL3-8B ↓1 83.4%
- 5
Qwen3-VL-8B-Instruct ↓1 82.4%
- 6 llava-v1.6-mistral-7b ↑2 78.1%
- 7
Qwen2.5-VL-7B ↓1 78.1%
- 8
Llama 3.2 11B ↓1 77%
- 9
InternVL3.5-8B 62.6%
Planning: SG
9 models we track have a score here.
- 1 GPT-5.4 88.3%
- 2
Qwen3-VL-235B-A22B 86.5%
- 3
InternVL3-8B 82.8%
- 4
Gemma 3 12B ↑1 82.2%
- 5
Qwen3-VL-8B-Instruct ↓1 78.5%
- 6
Qwen2.5-VL-7B 73.6%
- 7
Llama 3.2 11B 69.9%
- 8 llava-v1.6-mistral-7b 64.4%
- 9
InternVL3.5-8B 52.8%
Planning: IND
9 models we track have a score here.
- 1
Qwen3-VL-235B-A22B ↑1 91.4%
- 2 GPT-5.4 ↓1 90.1%
- 3
Gemma 3 12B ↑2 79.5%
- 4
Qwen3-VL-8B-Instruct 75.5%
- 5
InternVL3-8B ↓2 74.2%
- 6
Llama 3.2 11B ↑1 70.2%
- 7
Qwen2.5-VL-7B ↓1 60.9%
- 8 llava-v1.6-mistral-7b 53.6%
- 9
InternVL3.5-8B 53%
Region: CN
9 models we track have a score here.
- 1 GPT-5.4 99.4%
- 2
Qwen3-VL-235B-A22B 96%
- 3
InternVL3-8B 94%
- 4
Qwen3-VL-8B-Instruct 93.7%
- 5
Gemma 3 12B 82.1%
- 6
Qwen2.5-VL-7B 76.1%
- 7 llava-v1.6-mistral-7b ↑1 72.4%
- 8
Llama 3.2 11B ↓1 70.7%
- 9
InternVL3.5-8B 69.5%
Region: US
9 models we track have a score here.
- 1
InternVL3-8B ↑2 100%
- 2
Qwen3-VL-235B-A22B 100%
- 3 GPT-5.4 ↓2 100%
- 4
Qwen3-VL-8B-Instruct 97.7%
- 5
Gemma 3 12B 96%
- 6
Qwen2.5-VL-7B 91.5%
- 7 llava-v1.6-mistral-7b ↑1 87.5%
- 8
Llama 3.2 11B ↓1 86.9%
- 9
InternVL3.5-8B 69.9%
Region: UK
9 models we track have a score here.
- 1 GPT-5.4 100%
- 2
Qwen3-VL-235B-A22B 97.8%
- 3
Gemma 3 12B ↑2 90.6%
- 4
Qwen3-VL-8B-Instruct 87.7%
- 5 llava-v1.6-mistral-7b ↑3 78.3%
- 6
InternVL3-8B ↓3 76.1%
- 7
Qwen2.5-VL-7B ↓1 74.6%
- 8
Llama 3.2 11B ↓1 71%
- 9
InternVL3.5-8B 68.1%
Region: JP
9 models we track have a score here.
- 1 GPT-5.4 100%
- 2
Qwen3-VL-235B-A22B 85.8%
- 3
Gemma 3 12B ↑2 76.9%
- 4
Qwen3-VL-8B-Instruct 59%
- 5
Qwen2.5-VL-7B ↑1 57.5%
- 6
InternVL3-8B ↓3 55.2%
- 7
Llama 3.2 11B 50%
- 8 llava-v1.6-mistral-7b 45.5%
- 9
InternVL3.5-8B 24.6%
Region: SG
9 models we track have a score here.
- 1 GPT-5.4 99.2%
- 2
Qwen3-VL-235B-A22B 98.4%
- 3
Gemma 3 12B ↑2 91.1%
- 4
Qwen2.5-VL-7B ↑2 88.6%
- 5
InternVL3-8B ↓2 86.6%
- 6
Qwen3-VL-8B-Instruct ↓2 84.6%
- 7 llava-v1.6-mistral-7b ↑1 81.7%
- 8
Llama 3.2 11B ↓1 75.2%
- 9
InternVL3.5-8B 66.3%
Region: IND
9 models we track have a score here.
- 1 GPT-5.4 100%
- 2
Qwen3-VL-235B-A22B 99%
- 3
Gemma 3 12B ↑2 90.4%
- 4
InternVL3-8B ↓1 85.1%
- 5
Qwen3-VL-8B-Instruct ↓1 78.8%
- 6
Qwen2.5-VL-7B 76%
- 7 llava-v1.6-mistral-7b ↑1 71.6%
- 8
Llama 3.2 11B ↓1 65.9%
- 9
InternVL3.5-8B 51%
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.
Source: Official GeoDrive-Bench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.4 | Rule-Given | 91.9% | |||
| Rule-Given | 90.8% | ||||
| GPT-5.4 | Reasoning | 88% | |||
| Reasoning | 82.7% | ||||
| GPT-5.4 | Direct | 82.5% | |||
| Rule-Given | 81.3% | ||||
| Direct | 80.9% | ||||
| Rule-Given | 80.2% | ||||
| Rule-Given | 76% | ||||
| Rule-Given | 73.1% | ||||
| Direct | 70% | ||||
| Rule-Given | 68.9% | ||||
| Reasoning | 66.6% | ||||
| Direct | 65.3% | ||||
| Reasoning | 65.3% | ||||
| Direct | 64.8% | ||||
| Direct | 64.5% | ||||
| llava-v1.6-mistral-7b | Rule-Given | 62.8% | |||
| Reasoning | 56% | ||||
| Rule-Given | 53.8% | ||||
| Reasoning | 52.8% | ||||
| Direct | 51.2% | ||||
| llava-v1.6-mistral-7b | Reasoning | 47.3% | |||
| Reasoning | 44.4% | ||||
| llava-v1.6-mistral-7b | Direct | 41.7% | |||
| Direct | 38% | ||||
| Reasoning | 37.7% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official GeoDrive-Bench leaderboard
Qwen3-VL-235B-A22B on the official leaderboard passed 85% in September 2025
Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official GeoDrive-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 Robotics
- ERQA In use 400 multiple-choice visual questions from robot and egocentric footage on spatial, trajectory, action and state reasoning. Accuracy.
Gemini Robotics ER 2 78.5%
- SimplerEnv In use Real-to-sim replicas of Google Robot and WidowX/Bridge setups. Policies scored by success rate on pick, move, drawer and Bridge tasks. Praxis-1 80%
- RoboLab In use Zero-shot DROID Franka policies in Isaac Lab on 120 tabletop tasks at three instruction levels, 10 episodes each. Success rate. Praxis-1 46.2%
- RoboDojo In use Generalist manipulation policies run 42 Isaac Sim tasks and 18 remote real-robot tasks. Scored by task score and success rate. GPT-6 Astra 27.6%
- Progress classification (Google DeepMind) In use Models assign each frame of a robot video to one of five task-progress bands from 0 to 100 percent. Scored by accuracy.
Gemini Robotics ER 2 57.4%
- Human proximity (Google DeepMind robotics) In use Models judge from robot camera views whether a human is within 1 meter, a spatial awareness check for safe robot behavior. Scored by accuracy.
Gemini Robotics ER 2 93%
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 (9 entries).
- Pass Models scored 9 models we track.
- Note Newest score The newest came 8 months ago, in March 2026, for GPT-5.4. 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 91.9%, past 85% since September 2025.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 27 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.