Start with the key, not the bill
A refused key is one of those moments when a quick service starts to feel fiddly. Before you decide that SwiftRelay is difficult to use, check the simplest explanation: the key may be wrong for the model, the client, or the endpoint you are calling.
SwiftRelay is a pay-per-token API relay. You connect in minutes by copying three lines and running them. There is no monthly fee and nothing to subscribe to. You top up a balance, and each call deducts only what it used. New sign-ups receive $0.22 in gift credit, which gives you a small, practical way to make a first call before committing to a larger top-up.
That model is worth keeping in mind while troubleshooting. A refusal does not automatically mean you owe money or that the service is charging you for every possible attempt. It means the request did not reach a valid configuration. Start with the request, then move through the checks below in order.
- Confirm that the key you copied begins with swft_.
- Confirm that you are using the endpoint intended for your client.
- Confirm that the model belongs in that client.
- Confirm that the key is active and that your balance is not the problem.
Match the endpoint to the way you connect
One common source of doubt is the belief that every model should accept the same base URL and the same request format. SwiftRelay keeps the routes separate so that each client talks to the models it was designed to handle.
For Claude Code, use these three lines:
01 export ANTHROPIC_BASE_URL=https://api.fastapirelay.com 02 export ANTHROPIC_AUTH_TOKEN=<your swft_ key>
claude
For Codex, use these three lines:
01 export OPENAI_BASE_URL=https://api.fastapirelay.com/v1 02 export OPENAI_API_KEY=<your swft_ key>
codex
For Gemini CLI, use these three lines:
01 export GOOGLE_GEMINI_BASE_URL=https://api.fastapirelay.com 02 GEMINI_API_KEY=<your swft_ key>
gemini
For an OpenAI-compatible client, the base URL is https://api.fastapirelay.com/v1, and you enter your swft_ key as the API key.
Now the important detail: a key is not a universal pass. Claude models answer only through the official Claude Code or Claude Desktop client, with version 2.1.285 or later. Gemini models use only their native protocol through Gemini CLI. The OpenAI-compatible route pairs with GPT, DeepSeek and the other remaining models, never with Claude or Gemini.
So if a key is refused, check whether you combined a valid key with the wrong route. That mismatch is a configuration issue, not evidence that the key itself is faulty.
Check the client and model pairing
Here is the frank answer to the question a first-time reader is often too polite to ask: yes, the client matters, and yes, using the wrong one can stop a request before it becomes useful.
Claude models belong in the official Claude Code or Claude Desktop client, version 2.1.285 or later. If you are working with a Claude model in some other tool, the request is aimed at the wrong path. Gemini belongs in Gemini CLI and speaks its native protocol; it is not an OpenAI-compatible option. When you use the OpenAI-compatible endpoint, choose GPT, DeepSeek or another model that is available on that route.
This is why a short pairing check can save a long debugging session:
- Claude model: official Claude Code or Claude Desktop, version 2.1.285 or later.
- Gemini model: Gemini CLI using its native protocol.
- GPT, DeepSeek and other remaining models: the OpenAI-compatible endpoint at https://api.fastapirelay.com/v1.
Check the model identifier too, because a close-looking name is not the same model. SwiftRelay lists model names and per-call usage in the dashboard. You can also see current pricing on the SwiftRelay pricing page. For example, GPT-6.1 Sol is listed at $0.1736 per 1M input tokens, $0.8542 per 1M output tokens and $0.0347 per 1M cache-read tokens.
Those figures are useful when you compare the model you intended with the model actually being billed. They also make the pay-per-token idea concrete. If you are curious about the cost, an example is a single call using 1M input tokens and 1M output tokens with GPT-6.1 Sol: the example total would be $0.1736 plus $0.8542, or $1.0278. This is an example of arithmetic, not a quota or a fixed package.
Look for the small details that cause big refusals
Once the route and model are right, the problem is usually something small. A copied key can include a missing character, a stray space, or the wrong environment variable. If you paste a key into the variable meant for a different client, the tool may read an empty value or a stale one.
Run the three lines again in the shell or terminal where your client will actually start. Restart the client after changing environment variables. If you copied the key from the dashboard, compare it carefully with the value you pasted, especially the beginning and the final characters.
Next, check the account limits. SwiftRelay allows up to 20 live API keys per account. If you have reached that limit, a newly created key may not be the one in use, or the service may refuse another live key. Remove an unused live key if you need to make room.
Then check your balance and your recent calls. The usage view lists tokens and cost per call, so you can see whether a request is being accepted, rejected, or charged. Top-up amounts are $1.39, $6.94, $13.89, $27.78, $69.44 and $138.89. You can choose the amount that suits you rather than committing to a recurring arrangement.
If you top up and still see a refusal, note the time of the attempt and the model you requested, then open Messages in the dashboard. SwiftRelay support is handled through Messages, and login is required. There is no support email, so the dashboard is the place to take the question and the evidence together.
Know the refund and reward rules before you top up
Good value should come with clear terms, especially when you are new and trying a small amount first. SwiftRelay refunds unused credit within 11 days of a top-up, minus what was used and the payment fee. That gives you a defined window for testing the service without turning a trial into an open-ended commitment.
The balance model also keeps the arithmetic visible. There is no monthly fee and nothing to subscribe to. Each call deducts only what it used, and the usage list shows tokens and cost per call. If you want to compare this approach with keeping your own subscription or official account, you can do that in general terms: one pays for what you actually consume, while the other involves a commitment you continue whether or not you use it fully.
Sign-ups receive $0.22 in gift credit. SwiftRelay also offers referral rewards: 14% of each top-up by someone you invited, and 3% from the people they invite. Rewards can be withdrawn, with a minimum of $4.86, or moved into your balance.
None of that changes the first troubleshooting step. If a key is refused, verify the key, the endpoint, the client and the model pairing. If the request is valid but the answer is not what you expected, check the usage list and then use Messages in the dashboard.
Stopwatch out, keys checked, first run clean: SwiftRelay is ready when you are, so sign up now and make your first call.