Data Extraction Benchmark
The Data Extraction task asks each model to find and return one specific piece of information from an image, rather than all of its text. The target is usually a single field on a structured or semi-structured source: a price, date, timestamp, phone number, license plate, serial number, or meter reading. The model must locate the right field among many competing ones and return just that value in the requested format.
Evals updated August 27, 2026Pricing updated August 31, 2026
| 1 | 96.9% | 1.2K | $0.0008 | 1.98s | ||
| 2 | 94.8% | 1.3K | $0.0015 | 2.48s | ||
| 2 | 94.8% | 1.4K | $0.0037 | 2.67s | ||
| 2 | 94.8% | 1.5K | $0.0063 | 5.64s | ||
| 2 | 94.8% | 1.3K | $0.0014 | 8.73s | ||
| 6 | 92.8% | 1.5K | $0.015 | 4.98s | ||
| 7 | 90.7% | 1.2K | $0.0004 | 1.33s | ||
| 8 | 89.7% | 1.5K | $0.0030 | 2.81s | ||
| 9 | 88.7% | 1.5K | $0.0080 | 4.17s | ||
| 9 | 88.7% | 1.8K | $0.0033 | 4.18s | ||
| 11 | 87.6% | 1.5K | $0.0076 | 2.48s | ||
| 11 | 87.6% | 1.5K | $0.010 | 4.74s | ||
| 11 | 87.6% | 1.2K | $0.0029 | 5.59s | ||
| 14 | 86.6% | 1.6K | $0.0006 | 2.17s | ||
| 14 | 86.6% | 1.1K | $0.0002 | 2.85s | ||
| 14 | 86.6% | 1.1K | $0.0002 | 4.58s | ||
| 14 | 86.6% | 1.8K | $0.0031 | 4.63s | ||
| 18 | 85.6% | 1.1K | $0.0007 | 2.02s | ||
| 19 | 84.5% | 400 | $0.0013 | 3.02s | ||
| 19 | 84.5% | 1.6K | $0.0042 | 4.51s | ||
| 19 | 84.5% | 1.5K | $0.0046 | 5.65s | ||
| 22 | 83.5% | 1.1K | $0.0004 | 2.42s | ||
| 22 | 83.5% | 1.5K | $0.0001 | 4.78s | ||
| 22 | 83.5% | 1.8K | $0.0044 | 5.18s | ||
| 25 | 82.5% | 1.4K | $0.0014 | 2.84s | ||
| 25 | 82.5% | 1.4K | $0.0033 | 3.42s | ||
| 27 | 81.4% | 1.5K | $0.0004 | 2.94s | ||
| 27 | 81.4% | 1.4K | $0.0018 | 4.35s | ||
| 29 | 79.4% | 1.1K | $0.0005 | 1.24s | ||
| 29 | 79.4% | 1.4K | $0.0037 | 3.06s | ||
| 31 | 78.3% | 1.1K | $0.0002 | 2.20s | ||
| 31 | 78.3% | 1.1K | <$0.0001 | 3.49s | ||
| 31 | Kimi K2.6 | 78.3% | 1.6K | $0.0015 | 7.14s | |
| 34 | DeepSeek V4 Flash Vision Exp | 65.0% | 386 | $0.0002 | 1.26s |
Score vs. cost
Data Extraction score (Accuracy) against estimated cost per sample, on a log scale. Upper-left is the sweet spot: high quality at low cost.
34 models on the current benchmark · Data Extraction task only
Example Data Extraction benchmark tasks
Real samples from the benchmark: the image each model sees, the question it is asked, and the ground-truth answer it is scored against.
The models are asked
What is the license plate number on the rear of the black SUV? Format the answer as a continuous sequence of uppercase letters and digits (e.g., ABC1234).
Ground truth
9CGN302
The models are asked
What is the complete two-line identification text printed in white on the upper right side of the container doors? Format the answer with a single space separating the top sequence from the bottom sequence.
Ground truth
TCKU 612608 6 45G1
The models are asked
What is the large alphanumeric tire size sequence printed on the right side of the sidewall? Output the exact text including spaces and slashes.
Ground truth
205/60 R 16
How Data Extraction is scored
Each answer is graded against the ground-truth value. The leaderboard score is the percentage of samples answered correctly.
Every model runs the same sample set in a single evaluation pass. Token usage is measured from each provider’s API response, and cost per sample is that usage multiplied by the model’s published pricing. See the full methodology.
Frequently Asked Questions
OCR asks for a complete transcription of everything in the image. Data Extraction asks for exactly one value, so the model must find the right field among many competing ones and return it in the requested format. A model can be strong at full-page OCR and still fail at targeted extraction.
Each model answers the same extraction prompts. Answers are graded against the ground truth, and the score is the percentage answered correctly.
Most models in this leaderboard link to their Playground page. Click the model name to open it, then upload your own image and run it. A few models are benchmarked for comparison only and do not have a Playground page yet.