AI 原生开放教材

外观

AI Engineering Lab（24 周自学课程）系统课程·英文原文 ↗原件 ↓仓库 ↗

Skill: OpenRouter ​

Part of AI Engineering Lab · Developed by Zorost Intelligence AI Lab · zorost.com

1. What it is ​

OpenRouter is a unified LLM API and router: one account, one key, one OpenAI-compatible endpoint that reaches hundreds of models across many providers (Anthropic, OpenAI, DeepSeek, Meta, Mistral, Google, and more). It also adds automatic failover, so a provider error or timeout reroutes your request instead of killing the run.

It is not a model and not an editor, it's the routing layer behind a harness. Use it when you want to switch models by changing a slug instead of reconfiguring a provider.

Note. As of research time there is no product named "9Router", the closest match is OpenRouter (the phrase is almost certainly a mishearing). Treat "9Router" = OpenRouter.

2. Account & keys ​

  1. Create an account at openrouter.ai.
  2. Generate a key at openrouter.ai/keys.
  3. Call the API with a standard header:
bash
curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"meta-llama/llama-3.3-70b-instruct",
       "messages":[{"role":"user","content":"What is the predicted ETA for SH-4821?"}]}'

Keep the key in an environment variable, never in a prompt or committed file.

3. Routing one API to many models ​

Models are addressed by slug (provider/model):

anthropic/claude-sonnet-4
deepseek/deepseek-chat
meta-llama/llama-3.3-70b-instruct

Beyond plain slugs, OpenRouter adds:

  • Provider routing: pick specific providers, or let OpenRouter sort by price/latency/throughput.

  • Automatic fallbacks / model fallbacks: on error or timeout, reroute automatically.

  • Routers: higher-level helpers: Auto Router (pick a model for your prompt), Pareto Router (pick a coding model by minimum coding score), Fusion Router (multi-model deliberation as a slug), Free Models Router.

  • Variants: suffixes that change behavior: :free, :extended, :thinking, :online, :nitro, :exacto.

    VariantWhat it does
    :freeZero-cost, rate-limited access
    :thinkingExtended reasoning / thinking output
    :onlineAdds web-search grounding
    :nitroFaster, higher-tier routing
    :extendedLarger context variant
    :exactoDeterministic / exact-output behavior

  • BYOK: bring your own provider keys so you pay the provider directly (skips OpenRouter's token markup).

Switching models = editing one string. That's the whole point.

4. Model selection UI ​

Browse openrouter.ai/models to compare models by slug, context length, pricing per token (input/output/cache), and capabilities. The UI (and the /api/v1/models endpoint) is where you confirm a slug exists before wiring it into a harness, model names change often.

5. :free models ​

Append :free to a model id for zero-cost access:

meta-llama/llama-3.2-3b-instruct:free

Free variants carry stricter rate limits and availability than paid versions (per-model limits, commonly on the order of tens of requests per minute or a daily cap; check the rate-limits doc). Use them to test plumbing (auth, config, prompts) before paying, then swap to a paid slug for real throughput. Don't build a production loop on :free.

6. Cost tracking & limits ​

  • Pay-per-token against a prepaid credit balance; top up on the dashboard. You are charged input, output, and (when applicable) cache-read tokens at per-model rates.
  • OpenRouter adds a small markup for its routing/failover service; BYOK removes it if you already pay Anthropic/OpenAI directly.
  • Track spend in the dashboard (usage, per-model and per-key cost), and set usage limits and budget alerts so a runaway agent loop can't drain the balance.
  • Exact per-model prices are listed on each model's page, read them before quoting, and re-read them before a big batch, since they shift.

Quick cost playbook for a harness: use :free or a cheap small model to prove the config and prompt work → run real tasks on a mid-tier paid slug → reserve the expensive frontier models for the hard reasoning steps only → set a daily usage limit as a backstop.

7. Using it from your tools ​

  • OpenCode: add an openrouter provider with baseURL: https://openrouter.ai/api/v1 and set "model": "openrouter/<slug>" in opencode.json (see the OpenCode skill).
  • Claude Code / other OpenAI-compatible clients: point the base URL at OpenRouter and use your key; swap models by slug.
  • Ollama-compat / local: OpenRouter is a hosted endpoint, not a local one; for local models use Ollama/LM Studio. You can run the two side by side: :free hosted for one-off checks, Ollama for private/offline work.
  • Curl / SDKs: anything that speaks the OpenAI chat-completions shape works.

8. Pitfalls ​

Before relying on a new slug, run a one-line sanity check (auth + slug validity):

bash
curl -s https://openrouter.ai/api/v1/models | grep -o '"id":"[^"]*"' | grep "$SLUG"

If the slug isn't in the list, it's been renamed or retired, find its replacement before wiring it into a config.

  • Pricing changes. Per-token prices and markup shift without notice, re-check the model page before committing to a budget.
  • Model deprecation. Slugs get retired or renamed; a config that worked last month may return 404. Pin to a current slug and re-verify periodically.
  • :free is not for production. Rate limits and availability are much lower, treat it as a sandbox tier only.
  • Provider variability. The "same" model can behave differently across providers; if quality drifts, pin a specific provider rather than letting OpenRouter choose.
  • Secrets. The key is bearer-credential sensitive, keep it out of prompts, logs, and git.
  • Failover is not quality insurance. Automatic fallback keeps the request alive, but a fallback model may answer worse, don't treat "no error" as "good answer."
  • Cache-read billing. If your harness caches prompts, you may see cache-read token charges in addition to input/output, budget for all three line items.

9. Sources ​

Pricing, rate limits, and model slugs change frequently; re-check the docs before quoting.

© 2026 Zorost Intelligence LLC · https://zorost.com