OpenRouter raised $113M. LiteLLM does it for free. Pay anyway?

What happened

OpenRouter announced a $113M Series B on May 30, led by a16z and Menlo Ventures, valuing the company at roughly $1.5B. The round comes 14 months after a $40M Series A and follows a reported run-rate jump from ~$10M to over $100M in annualized revenue. The product is conceptually simple: a single OpenAI-compatible endpoint that fronts 300+ models from Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, Cohere, and dozens of inference providers including Together, Fireworks, DeepInfra, and Groq.

The pitch to developers hasn't changed since launch: one API key, one schema, one billing relationship. Swap `gpt-4o` for `claude-3.7-sonnet` or `deepseek-r1` by changing a string. OpenRouter takes a 5% markup on every token and, in exchange, handles provider auth, key rotation, fallback routing, and a normalized billing model that survives the weekly model-release churn.

The round itself isn't surprising — the multiples are. $1.5B for a company that owns no models, no GPUs, and no training data is the cleanest possible bet on the picks-and-shovels thesis. The interesting question isn't whether OpenRouter is valuable. It's whether the moat is defensible against the open-source stack that does literally the same thing.

Why it matters

There are at least four credible OSS alternatives shipping today. LiteLLM (BerriAI, ~16k stars) is the obvious one: a Python proxy with the same OpenAI-compatible schema, support for 100+ providers, cost tracking, fallback logic, and a self-hosted gateway. Portkey open-sources its gateway under MIT. OneAPI (the Chinese fork ecosystem) does multi-provider routing with a built-in admin UI. And vLLM plus Ollama cover the self-hosted inference path if you're running open weights directly.

The API-shape competition is over. Every router exposes `chat/completions`. The differentiator is what happens in the 50ms between the request hitting the proxy and the token stream returning. That's where OpenRouter has been quietly investing, and it's the only part of the stack that's genuinely hard to replicate.

Consider what real production routing actually requires. Provider A is 200ms faster but rate-limits at 60 RPM. Provider B has the cheapest tokens but drops one request in fifty during peak hours. Provider C just deployed a new model variant that's 30% cheaper but with subtly different tool-use semantics. A router that picks correctly across these dimensions, in real time, with sub-100ms decision latency, needs continuous benchmark data, per-provider health signals, and a learned cost-vs-latency policy. LiteLLM gives you the plumbing. OpenRouter gives you the policy.

The actual product is a constantly-updated leaderboard of provider behavior — uptime, p50/p99 latency, token throughput, refusal rates — fed back into the router as live state. This is what you're paying the 5% markup for. Not the abstraction. The telemetry corpus that makes the abstraction smart.

The community reaction on Hacker News (385 points, 200+ comments) was unusually split for an a16z fundraise. The skeptics pointed out, fairly, that markup-on-tokens businesses get squeezed the moment a major provider undercuts them — Anthropic and OpenAI both have first-party multi-model offerings via Bedrock and Azure that match price-to-price. The defenders made a different argument: developers don't want to write retry logic for the seven providers they actually use in production, and the 5% is rounding error compared to the engineering hours saved.

What this means for your stack

The decision is genuinely contingent on workload shape, and the framework is simpler than the marketing on either side suggests.

Use OpenRouter when: you're calling 3+ models across 2+ providers in production, you need automatic failover when Anthropic returns a 529 or OpenAI 429s, your team can't justify the on-call burden of a self-hosted proxy, or you're doing real-time model arbitrage (cheapest provider for a given prompt class). The 5% markup at $50k/month of inference spend is $2,500 — less than one senior engineer-week of routing maintenance.

Skip it when: you're single-model, single-provider (e.g., all-Claude or all-GPT-4o), you have predictable steady-state traffic that doesn't need elastic provider arbitrage, your latency budget can't tolerate an extra hop, or you're regulated and need direct vendor relationships for SOC2/DPA reasons. LiteLLM self-hosted gives you 90% of the value for the cost of a small EC2 instance.

The middle case: if you're running a small team, doing experimentation across models, and don't have prod traffic that justifies infrastructure investment, OpenRouter is the right default until you outgrow it. The migration path to self-hosted is straightforward because the API shape is portable. The lock-in everyone fears doesn't really exist — you can switch off OpenRouter in an afternoon, which is exactly why they have to keep earning the 5% every month.

Looking ahead

The interesting question for the next 12 months isn't whether OpenRouter survives — at $100M ARR with this round, they have runway through any plausible market shift. It's whether the foundation labs decide that owning the router layer matters enough to compete directly. Anthropic's MCP push and OpenAI's Assistants API both contain seeds of a vertically-integrated routing story. If either ships a credible cross-provider gateway with first-party economics, the 5% markup becomes a much harder sell. Until then, OpenRouter is the rare picks-and-shovels play where the picks actually got smarter than the alternative.