OpenRouter
Cloud-hosted model routing service that fronts hundreds of open and closed LLMs behind a unified API. Works with Claude Code via environment variables (ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN) in settings.local.json — point all model env vars (not just the default) at OpenRouter free models or cheap paid models to fully replace Anthropic billing. Includes a free router that load-balances across available free models.
Datadog’s State of AI Engineering report uses OpenRouter as an example of the managed gateway pattern teams adopt when direct API calls across many model providers become difficult to govern, benchmark, fail over, and optimize [src-037].
In Nate’s May 2026 stack video, OpenRouter is a specialist tool rather than a daily work surface. Nate says he left it out of the initial tier list by mistake but would place it in B tier for specialist routing use cases, parallel to key.ai for image/video models [src-053].
Key facts
- Integrates with Claude Code via env vars in .claude/settings.local.json
- Free models capped at 50 requests/day on accounts with no balance, 1,000/day with $10 loaded
- Must override ANTHROPIC_MODEL, HAIKU model, and SONNET model variables or Anthropic charges you for fallbacks
- Provides an openrouter/auto free router that selects the least-loaded free model
- Qwen 3.6 free with 1M context window available via OpenRouter
- Cheap paid models (Gemma 4 31B at ~14 cents / 40 cents per 1M tokens) can make Claude Code 50-100x cheaper than Opus
- Gateway services such as OpenRouter help teams route requests across model portfolios, enforce standards, and switch models as quality, latency, cost, or provider capacity changes [src-037]
Related concepts
- Model Fleet Governance
- LLM Capacity Engineering
- Practitioner Model Benchmarking Methodology
- Lean AI Tool Stack
Source references
- [src-004] Nate Herk cluster — Nate Herk — Claude Code cluster (21 videos)
– Videos referenced: O2k_qwZA8HU