Roboflow

OCR Benchmark

Gemini 4 Neon leads Roboflow's Vision Evals OCR benchmark at 88.1% (low effort, preliminary single run), with GPT-6.2 Sol at 87.5% and GPT-6.2 Astra at 87.4%. At high reasoning effort, Gemini 4 Carbon reaches 89.7%.

57 models evaluated114 runs across effort levelsPreliminary: single run

Evals updated November 19, 2026Pricing updated November 22, 2026

Score key:≥75%40–74%<40%
1high
89.7%
85.2%99.1%93.9%85.2%11.3K$0.10291.40s
2
GoogleGemini 4 NeonNEW
high
89.2%
83.5%98.8%93.9%86.0%6.5K$0.08249.87s
3
OpenAIGPT-6.2 SolNEW
high
89.0%
83.5%98.5%94.5%83.4%2.6K$0.01012.73s
4
OpenAIGPT-6.2 AstraNEW
high
88.3%
80.9%98.6%94.9%85.0%2.8K$0.04919.71s
5
GoogleGemini 4 NeonNEW
low
88.1%
80.9%98.5%93.8%86.3%2.4K$0.02120.95s
6high
88.0%
80.0%98.8%94.8%84.7%3.2K$0.02221.29s
7
OpenAIGPT-6.2 SolNEW
low
87.5%
80.0%98.1%94.2%83.6%2.2K$0.00598.15s
8
OpenAIGPT-6.2 AstraNEW
low
87.4%
80.4%98.8%93.7%82.7%2.2K$0.0248.73s
9high
87.2%
80.9%95.3%93.1%83.7%4.6K$0.02617.18s
10low
87.2%
80.4%96.6%93.0%84.2%1.5K$0.00537.54s
11low
86.7%
78.7%96.1%93.6%83.9%2.2K$0.008912.82s
12low
85.7%
78.3%95.5%92.6%81.5%1.5K$0.00517.70s
13
AnthropicClaude Fable 5.2NEW
high
84.8%
77.4%97.9%94.1%72.9%3.0K$0.04312.14s
14high
84.4%
74.8%97.0%93.4%79.2%2.9K$0.006618.75s
15low
81.9%
70.9%96.7%91.5%77.6%2.2K$0.003311.45s
16high
79.9%
69.6%96.1%94.1%61.1%3.1K$0.0238.96s
17low
79.7%
70.9%95.4%88.0%72.5%4.0K$0.007915.98s
18
AnthropicClaude Fable 5.2NEW
low
79.5%
64.8%94.7%92.8%74.5%2.7K$0.0309.03s
19high
78.8%
71.7%91.9%90.6%59.2%3.7K$0.01010.18s
20high
78.7%
70.9%94.1%90.9%58.6%10.5K$0.03979.76s
21low
78.3%
67.8%96.7%91.5%61.0%2.7K$0.0165.94s
22high
78.2%
66.5%94.2%89.2%71.3%4.1K$0.008417.40s
23low
77.8%
60.9%95.9%91.8%75.5%2.3K$0.03412.11s
24high
77.6%
59.6%97.5%91.7%76.8%3.3K$0.08439.33s
25high
77.4%
68.7%90.3%89.3%62.1%4.8K$0.01419.94s
26low
76.7%
71.7%90.0%86.8%56.2%2.5K$0.0137.39s
27high
76.0%
70.9%88.3%90.3%45.9%4.7K$0.05923.46s
28high
75.8%
56.5%96.0%90.4%76.7%3.1K$0.01629.89s
29low
75.8%
58.3%95.7%89.5%74.4%2.3K$0.006910.55s
30high
75.5%
69.6%85.4%88.1%53.3%4.5K$0.03114.36s
31high
75.3%
69.1%95.2%89.5%43.4%3.0K$0.05110.22s
32low
74.3%
67.4%93.2%90.5%40.1%3.4K$0.03412.30s
33high
74.2%
72.2%87.7%89.2%33.3%3.4K$0.07518.57s
34low
74.0%
63.9%95.2%85.4%59.0%2.3K$0.007718.12s
35low
73.9%
74.8%91.1%86.7%29.4%1.8K$0.00325.58s
36high
73.1%
68.7%90.8%83.9%47.1%3.2K$0.02618.64s
37high
73.1%
69.1%90.0%87.5%36.9%5.3K$0.01321.69s
38low
72.9%
64.3%84.7%89.2%46.5%2.7K$0.0397.94s
39low
72.8%
73.5%91.3%86.1%26.4%2.2K$0.004912.60s
40low
72.6%
67.8%88.5%82.0%52.0%1.9K$0.0108.51s
41low
72.3%
70.9%90.2%85.7%31.6%2.8K$0.0419.77s
42high
71.8%
61.7%94.8%81.6%59.6%5.7K$0.02878.42s
43low
71.8%
68.3%89.4%82.9%42.5%1.8K$0.00305.58s
44high
71.2%
66.1%88.6%82.1%46.4%3.6K$0.00639.60s
45
MetaMuse EmberNEW
high
70.5%
63.0%90.2%87.2%35.9%5.2K$0.00179.69s
46low
70.4%
62.6%91.8%88.0%33.3%3.6K$0.008626.68s
47high
69.8%
62.2%86.8%87.2%35.2%5.1K$0.01550.96s
48
MetaMuse EmberNEW
low
69.5%
61.7%95.5%85.1%35.3%3.7K$0.00106.66s
49low
68.8%
60.4%90.1%86.1%33.9%3.3K$0.00766.53s
50
OpenAIGPT-6 Sol
high
68.8%
47.4%91.0%85.3%68.7%3.4K$0.01826.43s
51low
68.7%
60.4%89.3%87.0%31.5%3.6K$0.008613.97s
52high
68.6%
61.7%85.1%86.7%30.7%4.9K$0.01426.58s
53high
68.6%
58.3%91.4%88.2%32.2%4.9K$0.01416.62s
54
AnthropicClaude Opus 5
low
68.0%
56.1%92.7%85.2%40.5%2.8K$0.0228.49s
55low
67.6%
60.0%89.0%85.5%29.5%3.1K$0.00615.83s
56high
67.5%
60.4%81.6%85.7%31.4%5.2K$0.01513.76s
57
OpenAIGPT-6 Sol
low
66.7%
45.6%88.3%83.4%65.7%2.4K$0.00779.27s
58low
66.3%
48.7%76.3%83.8%60.6%2.7K$0.00755.47s
59
AnthropicClaude Opus 5
high
65.8%
47.8%85.2%88.7%42.6%3.6K$0.04318.60s
60low
65.8%
64.3%87.5%75.8%31.9%1.6K$0.00124.65s
61high
65.5%
58.3%89.7%81.3%30.4%4.4K$0.004251.00s
62high
65.3%
46.5%94.7%85.3%45.8%4.0K$0.001012.27s
62high
65.3%
46.1%92.3%83.9%51.6%4.0K$0.04829.09s
64low
65.2%
60.9%86.8%76.3%35.9%825$0.00466.88s
65high
64.6%
63.0%87.3%78.3%20.6%2.8K$0.02521.02s
66low
63.9%
63.9%90.5%80.2%7.6%1.8K$0.00246.96s
67low
63.6%
54.4%89.5%86.4%14.8%5.7K$0.02338.94s
68low
63.6%
43.5%91.7%82.0%52.2%2.5K$0.01912.68s
69high
62.9%
51.3%87.3%79.3%37.4%2.7K$0.000815.40s
70high
62.5%
44.4%84.0%81.0%50.0%4.1K$0.001326.32s
71high
62.3%
56.1%83.8%75.7%31.7%4.5K$0.000445.77s
72high
62.1%
54.8%67.0%83.5%23.6%3.5K$0.03816.14s
73
MoonshotAIKimi K3
high
61.8%
43.9%80.4%82.9%43.4%6.2K$0.03336.90s
74high
61.7%
50.4%86.1%76.0%40.3%5.6K$0.01148.85s
75low
61.5%
53.0%70.8%82.4%24.2%3.0K$0.0259.48s
76low
61.2%
44.8%93.0%82.2%32.0%3.1K$0.00069.16s
77high
61.2%
50.9%88.7%85.7%9.4%6.5K$0.03070.49s
78low
60.7%
54.4%82.9%80.2%14.8%2.8K$0.001611.10s
79low
60.6%
50.0%89.0%84.4%11.0%3.2K$0.01118.68s
80
OpenAIGPT-5.5
high
60.6%
43.9%89.9%82.6%30.5%5.1K$0.10538.49s
81low
60.3%
53.5%86.7%75.8%23.1%1.8K$0.00097.87s
82high
60.1%
54.4%78.4%80.0%13.7%4.4K$0.004136.70s
83high
59.8%
49.1%80.9%74.9%36.8%4.4K$0.0097102.63s
84low
59.8%
41.7%80.0%78.2%47.4%2.5K$0.00049.31s
85high
59.5%
54.8%88.3%76.6%11.6%3.4K$0.01513.98s
86low
59.4%
52.2%88.6%78.3%13.1%2.7K$0.00786.54s
87low
58.9%
50.9%83.0%73.2%29.4%1.9K$0.00047.55s
88
MoonshotAIKimi K3
low
58.6%
43.0%75.1%78.1%39.3%4.5K$0.00536.35s
89
OpenAIGPT-5.5
low
58.4%
42.2%88.0%80.7%26.6%2.5K$0.02510.28s
90high
57.5%
47.8%86.1%80.9%7.0%3.4K$0.01322.48s
91low
57.1%
47.8%85.4%80.3%6.1%2.3K$0.006510.50s
92high
56.9%
37.4%80.2%77.3%41.9%3.4K$0.02121.67s
93low
56.4%
48.3%84.3%71.0%24.0%2.8K$0.001320.59s
94low
56.3%
36.1%79.9%76.8%42.3%2.6K$0.01110.54s
95high
56.0%
44.4%77.8%69.6%38.7%3.8K$0.008018.90s
96high
55.8%
47.0%78.5%79.0%6.1%7.4K$0.02733.34s
97high
55.7%
33.0%83.5%77.0%43.8%5.1K$0.004134.65s
98high
55.5%
46.5%84.1%71.4%21.6%2.8K$0.001321.86s
99low
55.4%
47.4%81.5%72.0%18.6%3.4K$0.000211.12s
100high
55.3%
50.9%72.2%72.0%14.5%4.2K$0.000224.14s
101high
54.5%
38.3%75.3%76.4%28.8%11.1K$0.021187.98s
102
MoonshotAIKimi K2.6
low
54.4%
34.8%77.9%78.5%29.8%2.5K$0.002610.55s
103low
54.3%
52.2%84.4%67.3%9.9%1.8K$0.00015.20s
104
MoonshotAIKimi K2.6
high
52.5%
38.7%54.2%75.8%26.9%8.1K$0.022107.65s
105low
51.8%
29.6%75.5%73.6%39.7%2.5K$0.00118.97s
106low
51.5%
43.0%78.0%67.5%17.7%2.7K$0.00395.55s
107low
51.1%
41.3%77.3%73.6%4.1%2.3K$0.00355.44s
108high
50.4%
42.2%68.9%69.2%13.2%5.5K$0.001259.84s
109high
49.9%
40.4%59.9%70.2%16.8%4.9K$0.003458.34s
110low
49.7%
41.7%72.5%68.6%9.2%2.4K$0.000422.21s
111low
48.6%
37.0%68.2%67.7%18.7%2.4K$0.00129.41s
112low
46.1%
33.5%68.9%71.7%0.0%3.2K$0.004448.82s
113
DeepSeekDeepSeek V4 Flash Vision Exp
high
44.7%
34.8%69.3%64.3%6.1%4.9K$0.006027.71s
114
DeepSeekDeepSeek V4 Flash Vision Exp
low
44.0%
34.4%65.2%64.9%3.4%5.1K$0.006230.11s

Score vs. cost

OCR score (Overall) against estimated cost per sample, on a log scale. Upper-left is the sweet spot: high quality at low cost.

57 models on the current benchmark · OCR task only, low effort

Example OCR 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.

Benchmark sample: Animal Ear Tag (Single Value)

The models are asked

Read the number on the animal’s ear tag. Return only the digits.

Ground truth

7072872

Benchmark sample: Analog Electricity Meter (Single Value)

The models are asked

Read the five-dial electricity meter. Return the five-digit reading, including any leading zeros.

Ground truth

99009

Benchmark sample: License Plate At Night (Single Value)

The models are asked

Read the largest visible vehicle license plate. Return only its registration number, without spaces; keep letters, digits, leading zeros, and hyphens.

Ground truth

KA52B3713

Benchmark sample: Comic Dialogue (Transcription)

The models are asked

Transcribe the speech and thought bubbles in reading order, one bubble per line. Preserve capitalization and punctuation; skip captions and sound effects.

Ground truth

All I'm asking for is a quick audience. Guess not. Look, all I'm saying is -- All right? I knew no one was listening.

Benchmark sample: Handwritten Italian Record (Transcription)

The models are asked

Transcribe all visible text in this Italian document in reading order. Preserve the original spelling, punctuation, and paragraph breaks; do not translate or include blank ruling lines.

Ground truth

ATTI DI NASCITA Giorgio, di anni quaranta, Segretaro, e Zanalda Davide di anni ventitre, Studente, entrambi residenti in questo Comune. La comparente ha fatto …

Benchmark sample: Words On A Tile Board (Transcription)

The models are asked

Read the valid Scrabble words formed by the tiles, including rotated words. Return them in uppercase, one per line, in alphabetical order; keep repeated words and ignore isolated letters.

Ground truth

CORDS DOOR JAILS JURY MUCK RAN SO SOT WITTY WREN

Benchmark sample: Shipping Container (Structured JSON)

The models are asked

Extract the main shipping container’s markings as JSON with these keys: container_id, size_type_code, maximum_gross_kg, maximum_gross_lb, tare_kg, tare_lb, payload_kg, payload_lb, capacity_m3, …

Ground truth

container_id: BMOU5848980; size_type_code: 45G1; maximum_gross_kg: 32500; maximum_gross_lb: 71650; tare_kg: 3850; tare_lb: 8490; payload_kg: 28650; payload_lb: …

Benchmark sample: Appliance Identification Label (Structured JSON)

The models are asked

Extract this appliance label as JSON with these keys: brand, model_number, product_number, serial_number, voltage, frequency, refrigerant, country_of_manufacture. Use strings, or null for missing …

Ground truth

brand: Electrolux; model_number: TR1171; product_number: 9244811-10; serial_number: 715-0410; voltage: 220 V; frequency: 50 Hz; refrigerant: R12; …

Benchmark sample: Sneaker Label (Structured JSON)

The models are asked

Extract this sneaker label as JSON: {"sku": null, "sizes": {"US": null, "UK": null, "EU": null, "CM": null}, "upc": null}. Use strings for visible values and null for missing fields. Prefer men’s …

Ground truth

sku: AI1407GLSC.X23

Benchmark sample: Circuit Labels (Text Localization) without annotations
Benchmark sample: Circuit Labels (Text Localization) with predicted boxes and labels
AnnotatedOriginal

The models are asked

Locate and transcribe each circuit label, value, and polarity mark, including rotated labels. Use tight boxes around individual markings and LaTeX for mathematical text; skip wires and circuit …

Ground truth

27 text regions: V_{DD}, +, M10, M11, V_{BP}, M6, …

Benchmark sample: Poster Typography (Text Localization) without annotations
Benchmark sample: Poster Typography (Text Localization) with predicted boxes and labels
AnnotatedOriginal

The models are asked

Locate and transcribe the poster’s text, including rotated or curved words. Use a tight box per text line or separate word group; skip decorative graphics. Return a JSON array of {"bbox": [x1, y1, …

Ground truth

2 text regions: FLOWER, POWER

Benchmark sample: Circuit Board Markings (Text Localization) without annotations
Benchmark sample: Circuit Board Markings (Text Localization) with predicted boxes and labels
AnnotatedOriginal

The models are asked

Locate and transcribe all visible markings on the circuit board and its components. Use one tight box per marking, join its lines with spaces, and skip logos and non-text symbols. Return a JSON array …

Ground truth

24 text regions: XP68-04A, R15, R16, U3, RM2, R50.0 D6L, …

How OCR is scored

Every question is scored deterministically against the ground truth, with no LLM judge. Single value answers must match exactly. Transcriptions are scored by character similarity, so a near miss earns partial credit. Structured answers are scored by how many JSON fields match, with missing and extra fields both counting against the model. Located text counts when a box overlaps the ground truth by more than half and its text matches exactly. The headline score is the mean over every question, and each category also has its own column. OCR results are currently a single preliminary run per model, so they carry no ±.

Runs, effort tiers and cost

Every model runs the same sample set. Under the current protocol each task is run three times at each effort tier and the score is the mean over those runs, shown with its ± range in the table; models benchmarked before the protocol ran once per task and are being re-run under it. 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

Each model answers the same questions. Every answer is scored from 0 to 100 against the ground truth by a rule that fits its category, and the headline score is the mean over all of them. Because the categories have different numbers of questions, the headline is not the plain average of the four category columns.

Single value: one reading returned exactly, such as a meter value or tire size. Transcription: all the text in a document or region, scored by similarity. Structured: the fields of a label, ticket, or form returned as JSON, scored field by field. Text localization: every piece of text with its bounding box, which tests whether the model knows where text is, not only what it says.

The OCR results come from one run per model at each effort tier, while the other tasks are averaged over three runs. A single run cannot show how much a score moves between runs, so these scores carry no ± range. They may be re-run three times later.

Every task is run at two levels of reasoning effort. The low row is the model answering with minimal deliberation, the high row is the same model on the same questions allowed to think longer. Both rows are ranked together so you can see whether the extra thinking is worth its cost and latency, which the Est. cost and Speed columns show, and the effort filter above the table shows one tier at a time. Models benchmarked before the current protocol have a high-effort row for Reasoning only until they are re-run. Every other table on this site, including the overall Average, uses the low-effort pass so all models are compared on the same amount of compute.

Under the current protocol each task is run three times at each effort tier, and the score shown is the mean over those runs. The ± is half the range between the lowest and highest run, so it shows how much the score moved between runs; it is a range, not a confidence interval. Hover the ± for the run count and range. A score without one comes from a single run; models benchmarked before the protocol are being re-run under it. OCR results are currently a single preliminary run per model, so they carry no ±.

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.