1
mistralai/mistral-small-2603
23.8232
WeirdBench
Benchmark leaderboard
nutrition-prediction
Nutrition Prediction
Predict calories, protein, carbs, and fat from ingredient lists for a fixed 50-dish Nutrition5k sample. Higher is better.
Higher score is better
2
openai/gpt-oss-20b
23.7467
3
z-ai/glm-5-turbo
23.5975
4
qwen/qwen3.5-122b-a10b
23.4188
5
mistralai/mistral-medium-3.1
22.9013
6
x-ai/grok-4.20-beta
22.5383
7
meta-llama/llama-4-maverick
22.0002
8
inception/mercury-2
21.8978
9
amazon/nova-lite-v1
21.6084
10
openai/gpt-5.4
21.5840
11
anthropic/claude-opus-4.1
21.5560
12
anthropic/claude-opus-4.7
21.5195
13
mistralai/mistral-large-2512
21.4432
14
openai/gpt-oss-120b
21.3529
15
openai/gpt-5.5
21.2108
16
amazon/nova-pro-v1
21.1849
17
openai/gpt-5.3-codex
21.1440
18
anthropic/claude-sonnet-4.6
20.7120
19
anthropic/claude-haiku-4.5
20.7097
20
openai/gpt-5.3-chat
20.6772
21
google/gemini-3.1-flash-lite-preview
20.5872
22
moonshotai/kimi-k2.5
20.1413
23
google/gemini-3-flash-preview
19.9842
24
openai/gpt-5.1
19.9581
25
google/gemma-4-26b-a4b-it
19.8426
26
google/gemma-4-31b-it
19.6220
27
anthropic/claude-opus-4.6
19.5945
28
z-ai/glm-5
19.4215
29
anthropic/claude-sonnet-4.5
19.3172
30
xiaomi/mimo-v2-pro
19.2432
31
moonshotai/kimi-k2.6
19.1032
32
anthropic/claude-opus-4.5
18.8760
33
x-ai/grok-4.1-fast
18.3064
34
stepfun/step-3.5-flash:free
17.9260
35
minimax/minimax-m2.7
17.5755
36
deepseek/deepseek-v3.2
17.4072
37
google/gemini-3.1-pro-preview
17.1302
38
meta-llama/llama-4-scout
16.6109
39
amazon/nova-2-lite-v1
16.0830
40
openai/gpt-5.4-mini
15.2246
41
amazon/nova-micro-v1
14.8388
42
minimax/minimax-m2.5
14.6662
Methodology
How scoring works
Fetch Nutrition5k dish metadata, deterministically sample 50 dishes that have at least 3 ingredients and 100+ calories, prompt the model once per dish, then compute per-field MAPE and Pearson correlation.
Prompt
Given only the ingredient list, return JSON with numeric `calories`, `protein`, `carbs`, and `fat` fields and no extra text.
Score
Higher is better. Overall score is 60% accuracy and 40% average correlation, where accuracy = 100 / (1 + average MAPE percentage).
Execution
Benchmark runners execute locally, use OpenRouter for predictions, fetch the fixed Nutrition5k metadata sample, cache results in Neon, and skip recomputation for models that already have stored scores.