OmniRoute/_ideia/_fetch/1812.json

{"author":{"id":"MDQ6VXNlcjE0NDAzMDM=","is_bot":false,"login":"matteoantoci","name":"Matteo Antoci"},"body":"## Problem\n\nIn `buildAutoCandidates()` (combo.ts), `costPer1MTokens` is set to only the input token price from `getPricingForModel()`. The output token price is read from pricing data but never used.\n\n## Details\n\nThis matters because many reasoning models (e.g., o3, DeepSeek R1) have cheap input tokens but expensive output tokens. Without factoring in output cost, these models appear artificially cheap in the scoring pool.\n\nFor example:\n- Model A: $3/M input, $15/M output → scored as $3\n- Model B: $5/M input, $5/M output → scored as $5\n- Router picks Model A as cheaper, but with a typical 40% output ratio, Model A actually costs $3 + $15×0.4 = $9 vs Model B's $5 + $5×0.4 = $7\n\n## Expected Behavior\n\nThe cost scoring factor should account for both input and output token pricing. A blended cost using an estimated output/input ratio would give more accurate cost comparisons between models.","comments":[],"createdAt":"2026-04-30T10:12:37Z","labels":[],"number":1812,"title":"[Feature] Auto-combo scoring ignores output token cost","url":"https://github.com/diegosouzapw/OmniRoute/issues/1812"}