Thinking Budget Testing

We tested the same specimen image with the same thinking level applied to both OCR and LLM parsing. The Jouvea Test measures whether OCR can read the complicated handwriting and whether the parsing model correctly identifies Jouvea pilosa as the specimen’s actual plant name. Distichlis spicata is also prominent on the label and is a tempting but incorrect choice; good models can resolve that ambiguity and pass this image. gemini-3.1-flash-lite on High is the recommended default and serves as the 1.00× base-cost case. gemini-3.7-flash on Low provided the best performance/cost tradeoff: it delivered very high performance and passed the Jouvea test for only a fraction more than the base case (1.42×, or 42% more). Higher thinking levels often consumed substantially more tokens and money without a proportional accuracy gain; the most expensive passing combination reached 41.35× the base cost. Google-hosted Gemma models receive low serving priority, can run very slowly, and produced server failures in this test, so we do not recommend them.

PASS means the returned formatted_json.scientificName contained “Jouvea” (case-insensitive). Costs reflect one specimen; “—” means no multiplier was available because no billable cost was recorded.