[MattFisher]

Summary

openrouter/<provider>/<model> names stopped resolving because LiteLLM dropped all of its OpenRouter rows (BerriAI/litellm#45665, 2026-10-09). The live service returns zero openrouter rows, and /api/inspect-costs returns 400 for every OpenRouter model. That leaves OpenRouter runs unpriced by the Inspect Costs Plugin.

This PR makes refresh also fetch OpenRouter's public /api/v1/models and convert each priced model into a LiteLLM-shaped row keyed openrouter/<id>. Those rows are merged into the cached dataset. The Inspect resolver needs no change: it already scopes openrouter/... lookups to litellm_provider == "openrouter" and matches the full key.

• Rates are OpenRouter's own. They can differ from the first-party provider's, and that is the right price for runs that go through OpenRouter. Its list also includes recent models LiteLLM lacks.
• Router models are skipped. openrouter/auto-beta and its siblings are priced at -1. 451 of 458 models are kept.
• One KV write. If either fetch fails, refresh throws and the previous cache stays in place. The trade-off: an OpenRouter outage also delays the LiteLLM refresh until the next six-hour run.
• Tiered prices are ignored. For example, Gemini's >200K-token prices are not used, so the base rate applies. This matches how LiteLLM rows are already exported.
• Capability flags (vision, tools, reasoning, prompt_caching) are derived from OpenRouter's fields, so the UI and API supports filters work on these rows.

Verification

• Unit tests cover the converter and resolution of openrouter/... names, including dotted versions such as claude-opus-4-1 against claude-opus-4.1. Worker tests cover the merged refresh and the failure path (previous cache kept).
• Against today's live LiteLLM and OpenRouter payloads, 13 OpenRouter judge models and 2 first-party controls all resolved to their own rows (for example, openrouter/openai/gpt-oss-20b: 0.018 input / 0.09 output per 1M tokens). First-party names still resolve to LiteLLM rows. The merged value is 2.2 MB, well under KV's limit.
• wrangler dev --test-scheduled ran the real refresh in workerd: 4,492 models cached, and /api/inspect-costs served the OpenRouter prices.
• npm test (31 passed), npm run typecheck, wrangler deploy --dry-run --env production, and pre-commit all pass.

Screenshots

Searching "openrouter" in the UI.

Before (production): 0 of 4041 models, and the footer credits only litellm.

After (local, refreshed): 451 OpenRouter rows with capability badges, and the footer credits both sources.

Not in scope

anthropic/claude-opus-4-1 (first-party) also fails to resolve. The cause is separate: LiteLLM's first-party anthropic rows now start at Claude 4.5, and 4.1 survives only as Bedrock and Databricks keys. openrouter/anthropic/claude-opus-4-1 resolves after this PR.

🤖 Generated with Claude Code