# Models, gateways, and cost ## Default `x-ai/grok-4.3` on OpenRouter (`OPENROUTER_API_KEY`) at `https://openrouter.ai/api/v1/chat/completions`. This default is fixed; do not substitute another model without an explicit reason recorded in the run notes. ## Order and cooldowns Try the gateway first, then OpenRouter. Each endpoint carries its own 60 second cooldown after a failure; a success clears it. Never call an endpoint inside its cooldown window; skip to the next one. ## Fallback tiers When the default is unavailable, fall through cheaper tiers in this order: `google/gemma-3-12b-it`, `google/gemma-3-12b-it:free`, `google/gemma-4-26b-a4b-it:free`, `google/gemma-4-26b-a4b-it`, `qwen/qwen3-30b-a3b`, `meta-llama/llama-4-scout`, `meta-llama/llama-4-maverick`. Free tiers first within equal capability; record which model produced what. ## Cost accounting Every gateway response contributes prompt tokens, completion tokens, and cost to running per-model totals. Unit prices per million tokens: gemma-3-12b 0.04 prompt and 0.13 completion, gemma-4-26b 0.08 and 0.35, qwen3-30b 0.08 and 0.28, llama-4-scout 0.08 and 0.30, llama-4-maverick 0.15 and 0.60. Sum into the job totals the host reads for budget enforcement. Never exceed the per-job spend cap; stop generating and write up partial findings instead.