OpenRouter: A Practical Guide

[ikaijua]

Introduction to OpenRouter (long-lived wiki page — contributions and corrections welcome).

1. What is OpenRouter?

The three-sentence version:

OpenRouter is an API gateway: with one API key and one bill, you can call 400+ AI models (OpenAI, Anthropic, Google, DeepSeek, Qwen, Kimi, Mistral, etc. across 60+ providers) through a single OpenAI-compatible API — existing code works by changing just the base URL.

What is an API gateway? A "switchboard" between your app and the model vendors. You tell it which model you want; it routes to the right provider, normalizes request formats, and tracks usage for billing. Switching models means changing one string, not re-integrating a dozen vendor APIs.

It trains no models and owns no GPUs — it only does the "model distribution" layer. By its own account it now processes 10+ trillion tokens per day, making it a real-world usage barometer (its model rankings are based on actual token volume, not lab benchmarks).

Company background: founded in 2023 by OpenSea co-founder Alex Atallah; raised a $113M Series B in May 2026 (led by CapitalG, ~$1.3B valuation); on August 19, 2026, Stripe announced its agreement to acquire OpenRouter for over $7 billion, with closing expected within weeks. Service continues as usual for now.

2. Quick start

1. Sign up at https://openrouter.ai (email or Google account)
2. Add credits on the Credits page (card purchases carry a 5.5% fee, $0.80 minimum; crypto 5%) — or skip this and start with free models
3. Create an API key on the Keys page
4. Point your code's base URL to https://openrouter.ai/api/v1 and use OpenRouter model IDs (format: vendor/model-name):

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="sk-or-...",
)
resp = client.chat.completions.create(
    model="anthropic/claude-sonnet-4.6",  # check the Models page for exact IDs
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

5. Verify: the call and its cost show up under Dashboard > Activity

The biggest gotcha: model IDs. Model names are not bare (gpt-5.6) — they are full vendor-prefixed IDs like openai/gpt-5.6 or deepseek/deepseek-v4.1-flash. A wrong ID fails with "model not found". Copy IDs from the Models page.

3. What it can do

• One key for every major model: OpenAI-compatible API across 400+ models / 60+ providers; switching models is a one-line change
• Smart routing and failover: popular models have multiple upstream providers; OpenRouter picks an available, low-latency one and fails over automatically
• Free models: 25+ models with a :free suffix cost nothing (rate-limited — see FAQ)
• BYOK (bring your own key): plug in your own vendor keys and pay no platform fee (up to a monthly allowance)
• Usage and cost dashboard: per-call model, token, and cost breakdown for accounting and limits
• Real-usage leaderboards: model rankings by actual token volume — closer to production reality than benchmark scores
• Native tool integration: many CLIs, IDEs, and agent tools have built-in OpenRouter support — paste a key and go

4. Typical use cases

• Give coding agents a cheap model: tools like Claude Code, Kimi Code, and OpenCode can connect to OpenRouter; route routine tasks to low-cost models like DeepSeek V4.1 Flash and cut spend by an order of magnitude
• A/B test models on the same code: loop over a few model IDs with the same prompts and compare quality and cost directly
• Add fallback in production: degrade automatically to a backup model when the primary is rate-limited or down — simpler than writing retries yourself
• Unify a team's model access: one shared account and balance; see who used what and how much, instead of managing a dozen vendor accounts per person

5. Advanced

• Provider routing preferences: pin or exclude specific upstream providers (e.g. official sources only)
• Prompt transforms: optional built-in prompt middleware, enabled per request
• Enterprise tier: higher BYOK allowance, SSO, etc., for team procurement
• Ecosystem signal: since mid-2026, Asian models (DeepSeek, Qwen, Kimi, and other open-weight models) have accounted for over half of OpenRouter's token volume — a window into the real open-vs-closed race

6. Pricing

Item	Details
Model usage	Billed at each provider's list price; OpenRouter adds no markup
Credit purchase fee	5.5% for cards ($0.80 minimum); 5% for crypto
BYOK platform fee	Free up to $25,000/month of list-price usage ($200,000 on Enterprise), then 5% above the allowance
Free models	25+ :free models, 20 RPM / 50 requests per day; 1,000/day once you have purchased $10 in credits (lifetime)
Account	Free to register; pay-as-you-go with no minimum spend

Rule of thumb: OpenRouter costs ~5% more than going direct (the purchase fee), in exchange for one key, failover, and consolidated billing. At ~$50K/month of model spend that 5% is real money, so it is worth keeping a self-hosted gateway (e.g. LiteLLM) as a plan B.

7. FAQ

Q: How is it different from calling the official API directly?
You pay roughly 5% more for one key across all models, automatic failover, and a single bill. If you only ever use one model and are latency-sensitive, go direct; if you switch models often or need fallback, OpenRouter saves real integration time.

Q: Is my data safe?
Requests pass through OpenRouter's servers (and, after closing, Stripe's orbit). Fine for general use; for sensitive or compliance-heavy workloads, call vendors directly or self-host a gateway.

Q: What about free credits?
No cash-style free credits, but 25+ :free models: 50 requests/day without purchase, 1,000/day once you have bought $10 of credits in total.

Q: Does it work from China?
OpenRouter is an overseas service; signup, payment, and API access all require working around network and payment barriers.

Q: Will the Stripe acquisition raise prices or shut it down?
As of September 2026, the official word is business as usual with unchanged pricing. Stripe's motivation is folding model-usage billing into payments infrastructure, so a shutdown seems unlikely — but heavy dependents should keep a self-hosted plan B.

---

Official references:

• OpenRouter
• Models
• API docs
• OpenRouter is Joining Stripe (official blog)
• Pricing