Open Gastronomy Standardv0.6

OGS-Bench

The benchmark of the Open Gastronomy Standard. Tasks put the OGS vocabularies and schemas in the prompt, so scores reflect reasoning with the standard rather than memorised trivia. Every task is scored deterministically and the aggregate is reported on a 0–100 scale.

Leaderboard Specification

Current standings.

Latest run per model, OGS-Bench v0.3. The leaderboard breaks every score down by category, language and task.

Benchmark Facts

Top 6 of 20 models evaluated

Top models by OGS-Bench aggregate score with 95% confidence half-width
ModelScore 0–100
1 gpt-5.586.8±3.2
2 claude-opus-4-785.9±3.6
3 gpt-5.485.3±3.6
4 gemini-2.5-pro85.0±3.0
5 gpt-4o80.9±3.6
6 gemini-2.5-flash80.8±3.8

213 tasks in 12 categories.

Category scores are means of their task scores; the aggregate weights them as set out in the specification.

Benchmark categories and task counts
CategoryMeasuresTasks
regionalCuisines, classic dishes, wine regions, grape varieties24
vocabularyKnowledge of the OGS controlled vocabularies23
pairingChoosing beverages for dishes22
sensoryTastes, structure and aromatics on the 0–10 scale21
compositionRoles, cooking methods, Maillard index, reduction20
explanationChoosing the right explanation codes20
riskSpotting negative pairing interactions20
structured_outputProducing valid OGS JSON from a brief20
dietaryDietary classifications and substitutions12
safetyFood-safety reasoning12
nutritionNutritional estimation10
menu_engineeringMenu balance, pricing and composition9

Every run is published in full.

Each run is a single JSON document with the task-set fingerprint, configuration, exact prompts, raw responses and per-task scores. All runs are published as they are. When a scorer defect is found, the stored responses are re-scored and the correction is recorded in the document itself.

To have a model evaluated, contact opensource@gastrona.app.