Vercel AI SDK on one API key

• A Node.js or Next.js project on the AI SDK (ai on npm).
• A RelaySwitch key. Sign in with an email code at the console and create one under Keys. Keys start with sk-relay-; the full key shows once at creation, so copy it then.
• A balance. Top-ups start at $10, by card or crypto, with no subscription and no KYC. Before the first top-up the console chat gives you 12 free turns a day to try models; SDK requests bill the balance.

1. Install the SDK core and the OpenAI-compatible provider package.
   bashCopy

   1pnpm add ai @ai-sdk/openai-compatible
2. Put the key in the environment, never in source. One key per app keeps the console Usage page readable.
   bashCopy

   1# .env.local (Next.js reads it; other runtimes use dotenv or the shell)

   2RELAYSWITCH_KEY=sk-relay-...
3. Build the provider once in a shared module and import it wherever you call a model.
   tsCopy

   1// lib/relayswitch.ts

   2// Options per the AI SDK OpenAI-compatible providers docs, read 2026-10-28:

   3// https://ai-sdk.dev/providers/openai-compatible-providers

   4import { createOpenAICompatible } from '@ai-sdk/openai-compatible';

   5

   6export const relayswitch = createOpenAICompatible({

   7 name: 'relayswitch',

   8 baseURL: 'https://api.relayswitch.io/v1',

   9 apiKey: process.env.RELAYSWITCH_KEY,

   10 // Adds token counts to streamed answers (stream_options.include_usage).

   11 includeUsage: true,

   12});

name labels the provider inside the SDK; it is also the key under providerOptions if you ever pass provider-specific options. baseURL keeps the /v1; the SDK appends /chat/completions. apiKey goes out as a Bearer token. Option names follow the OpenAI-compatible providers docs (read 2026-10-28).

relayswitch(id) returns a chat model, the same as relayswitch.chatModel(id). The id is the one the Models page shows, with no vendor prefix.

Streaming uses the same model object. With includeUsage on, the final chunk carries the token counts, so the result.usage promise resolves after the stream ends (streamText reference, read 2026-10-28).

Tool calling works the SDK way: each entry in tools is a tool() with a description, an inputSchema and an execute function, and stopWhen: stepCountIs(n) bounds a multi-step run (tool calling docs, read 2026-10-28). Every chat model answers the OpenAI chat completions format, and the Claude ids answer it as well as the Anthropic one, so the same provider covers Claude, GPT, Gemini, DeepSeek and GLM.

If @ai-sdk/openai is already in the project, it works here with the base URL set. Pick the Chat Completions factory:

A chat app resends the whole conversation every turn, so the input rate is most of the bill once a thread is long. The output rate matters for long answers and code generation. Of the models here, Claude Haiku 5.5 has the lowest input rate today at $0.0582 per million tokens.

Swapping models is one string: change the id in relayswitch(id). The full list, with every price and context size, is on the Models page. How the balance and the cached rate work is on Pricing.

The turn: 2,000 input tokens (a system prompt plus a short history) and 500 output tokens back. Those sizes are an assumption; the usage object on each result and the console Usage page show the real ones.

The same turn runs $0.00019 on Claude Haiku 5.5 and $0.0103 on Claude Opus 5.5. A cheap model for routine turns and a dear one for the hard question is a reasonable split; the provider block does not change.

Two curl calls rule out the key and the base URL before you debug the SDK. The first lists every model id your key can call:

The second sends one small request the way the SDK does. -i prints the headers, including the request id and, on a non-streamed answer, the charge in USD.

If both work, any failure left is in the provider block. Each SDK call then shows on the console's Usage page within seconds, with its tokens and its charge.

• 401: the key is wrong, revoked or unset. process.env.RELAYSWITCH_KEY is undefined when .env.local is missing or the variable name differs. Paste the key again from the console and run the curl check.
• 402: the balance is too low for the request. Top up in the console; the key works again at once.
• 404 on the model: the id is not in the catalog. Check for a typo or a display name in place of the id.
• 404 on the route: the base URL is missing /v1 or carries a trailing path. Use https://api.relayswitch.io/v1 exactly.
• 429: the key hit its rate limit. Back off and retry, or create a key with a higher Requests / min in the console.
• result.usage is empty on a stream: includeUsage is off. Set it to true in the provider block.
• 400 from @ai-sdk/anthropic on a non-Claude id: that package sends the Anthropic format to /v1/messages, which only the Claude ids answer. The OpenAI-compatible provider covers Claude and the rest with no second package.