TeX 59%
Python 40.8%
1# Extractor tiers. Model ids are exact API ids (Anthropic Claude 5.5 generation, verified via GET /v1/models on 2026-10-20).2# Thinking is kept minimal for a per-image classification task (cost control; thinking tokens are billed as output).3# - haiku 5.5: thinking disabled is accepted at effort <= high.4# - sonnet 5.5: thinking cannot be disabled; "between_tools" (no tools here → effectively off) at effort low.5# - opus 5.5: thinking cannot be disabled; effort low.6# Temperature is not set (Claude 5.5 models reject non-default sampling parameters when thinking is on).7tiers:8 haiku:9 id: claude-haiku-5-510 max_tokens: 120011 thinking: {type: disabled}12 effort: low13 price_per_mtok: {input: 0.10, output: 0.50} # list price; batches are billed at 50 %14 sonnet:15 id: claude-sonnet-5-516 max_tokens: 150017 thinking: {type: between_tools}18 effort: low19 price_per_mtok: {input: 2.00, output: 10.00}20 opus:21 id: claude-opus-5-522 max_tokens: 150023 effort: low24 price_per_mtok: {input: 4.00, output: 20.00}25batch_discount: 0.526