LiteLLM vs OpenRouter

When to pick LiteLLM

• You do not want a middleman on inference spend. The MIT gateway is free to self-host with no telemetry sent back to the vendor, you keep your own provider contracts and keys, and LiteLLM adds no token markup — OpenRouter’s 5.5% credit fee has a $0.80 minimum, so small top-ups pay an effective 10-20%.
• You are handing LLM access to an organisation. Virtual keys with per-key, per-user, per-team and per-org budgets, RPM/TPM rate limits, and tag-based cost attribution are all in the free build, which is the hard part of governing shared model access.
• Traffic must stay inside your network. Deployment covers Docker Compose, Kubernetes, Helm, and pip, with an air-gapped install available under the Enterprise licence, and local models sit behind the same OpenAI-format endpoint as hosted ones.

When to pick OpenRouter

• You want zero operations. There is nothing to deploy or upgrade, whereas running LiteLLM in production means owning Postgres, the proxy, and a regular upgrade cadence — only the four most recent stable minor lines get patches.
• Breadth and uptime matter more than fee percentage. OpenRouter fronts 400+ models from 70+ providers with automatic fallback when one is down or rate-limited, and only successful completions are billed.
• You want to know the price before you talk to anyone. Every self-serve rate is published, while LiteLLM’s Enterprise licence is quote-only and scoped to your deployment size — and SSO on LiteLLM is free only up to 5 users, with SCIM, JWT auth, and audit logs behind the paid licence.

Bottom line

These solve the same problem from opposite ends: LiteLLM gives you a gateway you run, OpenRouter gives you a gateway someone else runs. For a team with platform engineers and existing provider contracts, LiteLLM is close to free and strictly better on control — budgets, spend attribution, and rate limits in the MIT build, over 54,000 GitHub stars, and a roughly weekly release cadence. For a small team, a prototype, or anyone who wants one integration to cover 400+ models with failover, OpenRouter’s 5.5% credit fee is cheap next to the cost of operating a proxy. Two limits deserve attention: LiteLLM is not fully open source — everything under its enterprise directory carries a separate licence that forbids production use without a subscription, including several prompt-injection guardrail callbacks — and OpenRouter’s free models are capped at 50 requests a day, which makes them a testing surface rather than a production tier.