Put OpenRouter in front of production traffic only after you have tested the hosts behind each model and capped every key. Separate environments into workspaces, allowlist tested hosts with a guardrail, set data policy, and reconcile the fee, failed requests, and cache hits every month. This checklist covers hosts, keys, data policy, sign-in, failures, and cost.
When you need this
- An app or agent is moving from OpenRouter in development to OpenRouter in production.
- Several teams share one OpenRouter organization and nobody can say which key spends what.
- Security asks where prompts go, who can change settings, and how a leaked key is contained.
Evaluation criteria
Test the hosts behind each model
Which hosts are allowed to serve each production model?
Look forCompare hosts on the per-host benchmark closest to your workload and on your own prompts. OpenRouter's per-host benchmarks, as Olly's founder reports, put DeepSeek V4 Flash at 90% GPQA and 81% TAU-Bench at DeepSeek against 75% and 58% at DigitalOcean on the same weights, and found declared precision did not predict quality. Allowlist the hosts that pass with a guardrail, and keep fallbacks on within that list.
Separate environments and cap keys
Who holds which key, and how much can it spend?
Look forGive each environment its own workspace: Free and Standard allow 5, Business 1,000. Give each person, app, or agent a key with a credit limit, set a maximum key lifetime, and review keys with no limit on the Security settings page. One key with a US$10 daily limit spent US$100 in a burst, so do not treat limits as the only control.
Set data policy
Where may prompts go, and who may keep them?
Look forEnable Zero Data Retention account-wide or per request, and use eu.openrouter.ai or us.openrouter.ai on Business or Enterprise for in-region routing. Leave input and output logging off unless you need stored prompts. Batch API inputs and results are kept 30 days unless deleted, and a Private Endpoint retains whatever you declare for it.
Decide how people sign in
Do you need SSO, SCIM, or an audit trail?
Look forSSO (SAML), SCIM, contractual SLAs, and editing the IP allowlist for keys are Enterprise features. OpenRouter's audit log covers changes made through SCIM group mappings, so record guardrail and key changes in your own change log and use the Logs page's member column to attribute requests.
Plan for failures
What happens when a host errors or rate limits?
Look forOpenRouter retries other hosts on errors and 429s, and an attempt that errors is not charged for its tokens. Clients still see some failures as finished responses: a Pi issue reports a finish reason of error ending an agent turn, and practitioners report empty 200 responses. Retry those in your client, and alert on them from the Logs page.
Reconcile cost
What does a month really cost?
Look forAdd the 5.5% Standard or 8% Business fee on credit purchases to the model bill, or move steady traffic to BYOK, which carries no fee up to US$25,000 of list-price inference a month and 5% after. Check cache hits per model: one production game found about 1% before adding cache breakpoints.
Questions to ask vendors
- Which hosts have we tested for each production model, and when did we last retest them?
- Which workspace and key does each app and agent use, and what is its credit limit?
- Do we need Enterprise for SSO, SCIM, IP allowlists, or an SLA?
- Who reads the Logs page each week for empty completions, fallbacks, and cache misses?
- What changes for us if Stripe's acquisition closes and fees or terms move?
Red flags
- Production traffic routed to any host OpenRouter picks, with no allowlist.
- A host pin with fallbacks turned off and no second host.
- Keys with no credit limit and no expiry shared across apps.
- Prompt caching assumed in the budget but never checked in the Logs page.
- A security review that expects an audit log of every admin change.
Products in this piece
Sources
- Official documentation
- 12
- Public source
- 5
- OpenRouter docs: Provider routing
- OpenRouter docs: Model fallbacks
- OpenRouter docs: Guardrails
- OpenRouter docs: Security settings
- OpenRouter docs: Zero Data Retention
- OpenRouter pricing and plan comparison
- OpenRouter docs: Bring your own keys
- OpenRouter docs: Logs
- OpenRouter Enterprise
- OpenRouter documentation index (llms.txt)
- OpenRouter blog: Batch API (Sep 22, 2026)
- OpenRouter blog: Private Endpoints (Oct 14, 2026)
- Mo Moustafa: So you want to use OpenRouter? (Sep 7, 2026)
- Hacker News comment on a key with a US$10 daily limit spending US$100 (Sep 11, 2026)
- Garblesnarff/infinite-realms-app issue #633: DM model audit with a 1% prompt cache hit rate (Oct 14, 2026)
- earendil-works/pi issue #9002: OpenRouter finish reason error is never retried (Sep 2, 2026)
- Arseny Zinchenko, RTFM: LiteLLM and OpenRouter integration and fallbacks configuration (Jul 31, 2026)