Skip to main content
dbb.

Results

We test how accurately AI models identify dog breeds from photos, measuring first-guess and top-three accuracy, response time, and estimated cost.

500 photos across 25 dog breeds. 10 model configurations.

First match

Higher is better

How often the model’s first breed guess is correct.

  1. Claude Opus 5.5Medium reasoning86.0%
  2. GPT-6.1 TerraMedium reasoning83.8%
  3. GPT-6 AstraMedium reasoning83.6%
  4. GPT-6.1 SolMedium reasoning83.0%
  5. GPT-5 MiniMedium reasoning82.0%
  6. Claude Sonnet 5.5Medium reasoning81.0%
  7. Claude Haiku 5.5Medium reasoning80.8%
  8. GPT-4o MiniNo reasoning setting79.0%
  9. GPT-6 LunaMedium reasoning72.2%
  10. GPT-6 Luna FastMedium reasoning71.4%

Top three

Higher is better

How often the correct breed appears among the model’s three guesses.

  1. GPT-6 AstraMedium reasoning95.6%
  2. GPT-6.1 SolMedium reasoning95.2%
  3. Claude Opus 5.5Medium reasoning95.0%
  4. GPT-6.1 TerraMedium reasoning94.4%
  5. GPT-5 MiniMedium reasoning92.8%
  6. GPT-6 LunaMedium reasoning91.4%
  7. GPT-6 Luna FastMedium reasoning91.0%
  8. Claude Haiku 5.5Medium reasoning90.6%
  9. Claude Sonnet 5.5Medium reasoning89.4%
  10. GPT-4o MiniNo reasoning setting89.0%

Median latency

Lower is better

  1. Claude Haiku 5.5Medium reasoning1.58 s
  2. GPT-4o MiniNo reasoning setting2.01 s
  3. Claude Sonnet 5.5Medium reasoning2.21 s
  4. GPT-6 Luna FastMedium reasoning2.65 s
  5. Claude Opus 5.5Medium reasoning3.69 s
  6. GPT-6.1 TerraMedium reasoning4.66 s
  7. GPT-6 LunaMedium reasoning5.26 s
  8. GPT-6.1 SolMedium reasoning7.99 s
  9. GPT-6 AstraMedium reasoning9.43 s
  10. GPT-5 MiniMedium reasoning12.77 s

Est. cost / 1,000 scans

Lower is better

  1. GPT-6 LunaMedium reasoning$0.36
  2. Claude Haiku 5.5Medium reasoning$0.37
  3. GPT-6 Luna FastMedium reasoning$0.71
  4. GPT-4o MiniNo reasoning setting$1.54
  5. GPT-5 MiniMedium reasoning$2.69
  6. GPT-6.1 TerraMedium reasoning$2.74
  7. GPT-6.1 SolMedium reasoning$5.77
  8. Claude Sonnet 5.5Medium reasoning$7.43
  9. Claude Opus 5.5Medium reasoning$14.84
  10. GPT-6 AstraMedium reasoning$28.46

Median latency is measured per request. Serving conditions varied between runs. Some runs used prompt caching. Cost estimates use uncached model token rates for breed-analysis attempts, in USD at the recorded list rates.

Accuracy by breed

Each breed has 20 photos.

Average gives each of the 10 models equal weight.

First match accuracy by breed
BreedAverageOpusTerraAstraSol5 MiniSonnetHaiku4o MiniLunaLuna Fast
Soft-coated Wheaten Terrier
81.0%10 models
100%20/20
85%17/20
95%19/20
95%19/20
95%19/20
40%8/20
95%19/20
80%16/20
70%14/20
55%11/20
American Bulldog
61.5%10 models
60%12/20
95%19/20
70%14/20
65%13/20
40%8/20
70%14/20
50%10/20
70%14/20
50%10/20
45%9/20
American Pit Bull Terrier
45.0%10 models
35%7/20
45%9/20
45%9/20
50%10/20
75%15/20
45%9/20
40%8/20
65%13/20
30%6/20
20%4/20
English Cocker Spaniel
62.0%10 models
80%16/20
75%15/20
75%15/20
80%16/20
55%11/20
35%7/20
45%9/20
25%5/20
70%14/20
80%16/20
English Setter
78.5%10 models
95%19/20
90%18/20
85%17/20
85%17/20
90%18/20
85%17/20
90%18/20
75%15/20
50%10/20
40%8/20
Miniature Pinscher
83.0%10 models
100%20/20
50%10/20
85%17/20
90%18/20
100%20/20
100%20/20
95%19/20
100%20/20
60%12/20
50%10/20
Staffordshire Bull Terrier
54.0%10 models
65%13/20
60%12/20
75%15/20
80%16/20
55%11/20
60%12/20
45%9/20
40%8/20
30%6/20
30%6/20
Leonberger
87.0%10 models
85%17/20
100%20/20
95%19/20
100%20/20
85%17/20
80%16/20
95%19/20
85%17/20
60%12/20
85%17/20
Havanese
59.0%10 models
75%15/20
75%15/20
45%9/20
45%9/20
70%14/20
65%13/20
60%12/20
55%11/20
50%10/20
50%10/20
Japanese Chin
84.0%10 models
90%18/20
90%18/20
90%18/20
85%17/20
90%18/20
85%17/20
80%16/20
60%12/20
85%17/20
85%17/20
Keeshond
90.0%10 models
100%20/20
95%19/20
95%19/20
95%19/20
95%19/20
95%19/20
90%18/20
95%19/20
70%14/20
70%14/20
Scottish Terrier
62.0%10 models
70%14/20
65%13/20
75%15/20
70%14/20
45%9/20
75%15/20
65%13/20
60%12/20
50%10/20
45%9/20
Boxer
87.5%10 models
95%19/20
85%17/20
85%17/20
85%17/20
85%17/20
95%19/20
100%20/20
90%18/20
75%15/20
80%16/20
Chihuahua
80.5%10 models
85%17/20
85%17/20
75%15/20
65%13/20
75%15/20
85%17/20
80%16/20
90%18/20
85%17/20
80%16/20
Newfoundland
82.5%10 models
75%15/20
80%16/20
80%16/20
80%16/20
75%15/20
95%19/20
85%17/20
70%14/20
95%19/20
90%18/20
Pomeranian
93.0%10 models
95%19/20
100%20/20
90%18/20
90%18/20
95%19/20
95%19/20
75%15/20
100%20/20
95%19/20
95%19/20
Basset Hound
91.5%10 models
100%20/20
95%19/20
100%20/20
90%18/20
90%18/20
80%16/20
95%19/20
80%16/20
95%19/20
90%18/20
Great Pyrenees
94.5%10 models
100%20/20
95%19/20
90%18/20
90%18/20
100%20/20
100%20/20
100%20/20
80%16/20
95%19/20
95%19/20
Samoyed
89.5%10 models
95%19/20
90%18/20
95%19/20
95%19/20
90%18/20
95%19/20
85%17/20
95%19/20
75%15/20
80%16/20
Shiba Inu
91.0%10 models
95%19/20
90%18/20
85%17/20
85%17/20
100%20/20
95%19/20
95%19/20
100%20/20
85%17/20
80%16/20
Yorkshire Terrier
87.5%10 models
90%18/20
95%19/20
90%18/20
85%17/20
85%17/20
90%18/20
95%19/20
95%19/20
75%15/20
75%15/20
Beagle
93.0%10 models
95%19/20
90%18/20
90%18/20
95%19/20
95%19/20
95%19/20
90%18/20
100%20/20
90%18/20
90%18/20
German Shorthaired Pointer
84.5%10 models
85%17/20
85%17/20
90%18/20
90%18/20
80%16/20
85%17/20
80%16/20
80%16/20
85%17/20
85%17/20
Pug
98.0%10 models
95%19/20
95%19/20
100%20/20
100%20/20
100%20/20
95%19/20
100%20/20
100%20/20
95%19/20
100%20/20
Saint Bernard
87.0%10 models
90%18/20
85%17/20
90%18/20
85%17/20
85%17/20
85%17/20
90%18/20
85%17/20
85%17/20
90%18/20

Methodology

Each model evaluated the same 500 photos once with the same benchmark prompt. A match requires a valid breed ranking with the reference breed first or among the top three. All attempts count.

Dataset credit: Oxford-IIIT Pet dataset (Parkhi, Vedaldi, Zisserman, and Jawahar, 2012), CC BY-SA 4.0. Photo copyrights remain with their owners.