You don't reach a model one way. You reach it through a channel — and the channel is its price. The same call can be free on one and metered on another.
A model that had been working through its own direct key suddenly stopped — not an outage, not a code change. The account behind that channel got suspended, and every call through it died with it. The model still existed; the door to it was shut.
"I thought 'I have access to the model' was one thing. It isn't. The model was fine — the way I was paying to reach it was the thing that broke."
"Right — 'access' is really 'which channel.' Same Kimi, but the direct Moonshot key was suspended, so we re-routed it through OpenRouter and it came back. The model and the channel are two separate things, and the channel is the one that controls the bill and the on/off switch."
There are three ways to reach a model, and which one you use is what it costs you. Same model, three different bills (and three different on/off switches). Knowing the map is how you control spend instead of being surprised by it.
metered, per tokenPay-per-use, direct from the provider. You're billed for every token in and out. Maximum control, a live meter.
flat monthlyClaude Max/Pro, ChatGPT — included usage for a fixed fee. The call feels "free" because you already paid.
one balance, many modelsAn aggregator: one prepaid balance reaches many providers' models through a single endpoint. No new account per model.
Why it matters: the same Claude call is "free" on the subscription but metered on an API key — identical work, different bill. An obscure model you'd otherwise need a whole new account for is one OpenRouter line away. And a channel can be cut (a suspended key) without the model going anywhere — so knowing the map is also knowing your fallback.
You run the exact same Claude request two ways. Through your subscription it costs you nothing extra; through an API key it shows a per-token charge. Why?
subscription — included usage, no per-call bill while you build.API key — you pay per token, so you watch the meter.OpenRouter balance — one endpoint, many models.Same rule as always: answer it right, or the module isn't done.
You don't pick "AI" and you don't just pick a model — you pick a channel. The channel is the price, and the thing that can be switched off.
A real reference card — the three channels side by side (what they cost, when to reach for each, the gotchas), grounded in the actual channel map: Claude Code on the subscription, provider keys metered, Kimi/DeepSeek via OpenRouter. Paired with a fill-in MODEL-CHANNELS.md so you keep your own decision-of-record: every model you use, which channel it's on, and its fallback if that channel goes dark.