Fix model pricing by selecting best provider and deriving cache prices

[GoJam11]

Summary

This PR fixes a critical issue where model pricing data was being incorrectly overwritten by secondary resellers, resulting in missing cache pricing information. The fix implements intelligent provider selection, cache price backfilling, and fallback pricing derivation across the entire pricing pipeline.

Root Cause

The models.dev API exposes the same model ID (e.g., claude-opus-4-6) across 170+ providers. The previous implementation simply overwrote pricing with each provider encountered, meaning the final pricing came from whichever provider was last in the iteration order. This frequently resulted in secondary resellers' incomplete pricing (missing cache_read/cache_write fields) overwriting official provider data, causing cache operations to be billed at $0.

Key Changes

Model Catalog (src/model-catalog.ts)

• New buildCatalogMapsFromModelsDev() function: Replaces simple overwriting with intelligent provider selection
  • Ranks providers by trustworthiness (official/first-party providers prioritized)
  • Scores candidates based on pricing completeness and cache field availability
  • Selects the best candidate per model ID
  • Backfills missing cache_read/cache_write from other providers with identical base pricing
  • Validates all numeric fields (rejects negative values, non-numbers, missing required fields)

Pricing Calculation (src/pricing.ts)

• New resolveEffectiveCachePricing() function: Derives cache prices when upstream data is incomplete
  • Anthropic: cache_read = 0.1× input, 5m cache_write = 1.25× input, 1h cache_write = 2× input
  • OpenAI: cached_input = 0.1× input, no separate cache write fee
  • Marks derived prices with cache_pricing_derived flag for transparency
• TTL-aware cache write billing: Splits cache_creation_input_tokens into 5m and 1h buckets using new ephemeral_5m_input_tokens/ephemeral_1h_input_tokens fields
• Enhanced CostBreakdown: Adds cache_write_5m_tokens, cache_write_1h_tokens, and actual billing prices (cache_read_price, cache_write_5m_price, cache_write_1h_price)

Database (src/catalog-db.ts)

• Fixed onConflictDoUpdate to reference excluded.* (new values) instead of existing columns, ensuring refresh updates actually persist

Console Frontend

• Models page: Added cache pricing columns with improved formatting (preserves significant digits for sub-$1 prices)
• Cost breakdown display: Shows actual billing prices (with "(推导)" suffix for derived prices), splits 1h cache writes into separate formula rows
• Usage aggregation: Tracks ephemeral_5m/1h fields for consistent cost reporting

Impact

For a typical high-cache-hit request with claude-opus-4-6:

• Before: $0.000385 (only input/output tokens billed, cache tokens at $0)
• After: $0.033009 (correct billing: input $0.000010 + output $0.000375 + cache_read $0.027387 + cache_write $0.005237)

Affected ~1,606 of 2,652 models that were falling back to incomplete secondary reseller pricing.

Testing

Added comprehensive test suite (test/model-catalog-pricing.test.ts) covering:

• First-party provider preference over resellers
• Cache price backfilling from same-price providers
• Rejection of mismatched base prices
• Free/placeholder entry handling
• Malformed data validation
• Context window fallback behavior

https://claude.ai/code/session_011a12zJk3jbzU6orYUtiD4X