How can we help?

Sign up — new accounts get 1.3M+ free tokens. Your API key appears in the dashboard as soon as your account has a balance. Copy it right away: it is shown only once.

OpenAI-compatible SDKs and tools use https://api.cheaptokens.dev/v1. Anthropic-compatible clients, including Claude Code, use the root address https://api.cheaptokens.dev and add the path themselves. If your client adds /v1 on its own, give it the root address too.

Put your key in the CHEAPTOKENS_API_KEY environment variable and run the request below. The quickstart builds the same snippet for Python, Node and the Anthropic SDK, with the model of your choice.

curl https://api.cheaptokens.dev/v1/chat/completions \
  -H "Authorization: Bearer $CHEAPTOKENS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3-Flash",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Any tool that speaks the OpenAI or Anthropic API works: set the base URL, paste your key and add the model ID. Each guide shows where these settings are in the tool.

All models work with the same key and the same address — to switch, change the model ID. Each card on the models page lists the context window, maximum output and price.

• GLM-5.3-Flash For complex coding & reasoning
• DeepSeek-V4-Flash For everyday tasks on a budget
• MiniMax-M2.7 For agents & workflow automation

There is no subscription. You top up a USD balance, and each request is charged at the model's fixed rate per million input and output tokens. One balance covers every model.

In the dashboard, press Top up and choose the amount in USD. Pay with crypto (no KYC) or by card through a partner — the funds are credited to your balance automatically.

A crypto top-up is credited automatically once the network confirms the transaction — how long that takes depends on the coin. If you pay by card through a partner, the balance is topped up 5–30 minutes after payment.

Check

The status on the payment page: the full amount is paid, in the coin and network the page shows.

Fix

Wait — there's nothing to resend. You can close the payment page: the top-up is credited without it.

Make sure it works

The balance in the dashboard grows by the top-up; after your first top-up the API key appears there too.

Still failing? Send us

If the payment page shows the payment as complete but the balance hasn't changed within 30 minutes — see the next answer.

Payments are matched to your account automatically. If the notification about a crypto payment was lost, a regular check picks the payment up within minutes.

Check

The payment page or your wallet: the payment is complete. A crypto invoice must receive the full amount — a payment that arrived short is checked by hand. If you pay by card through a partner, the partner must send USDT on the TRC20 (Tron) network to your recipient address: other coins and networks are not credited.

Fix

Give it 30 minutes after the payment is complete. If the balance still hasn't changed, write to us — don't pay again.

Make sure it works

The dashboard shows the top-up in your balance.

Still failing? Send us

The payment method, the amount, when you paid and the payment ID or transaction hash from the payment page or your wallet.

Most often the amount is below the minimum — $5 with crypto, $10 by card through a partner — or the payment service didn't respond for a moment.

Check

The amount and the exact error text in the top-up window.

Fix

Reload the dashboard and try again in a few minutes. If paying by card through a partner fails, the top-up window offers a second partner.

Make sure it works

The payment page opens with your amount; after payment the balance grows.

Still failing? Send us

The payment method, the amount, the time and the exact error text — a screenshot helps.

No. Unused balance does not expire, so you can top up once and spend it at your own pace.

Requests return 402 and your key is paused. Top up, and the same key works again — you don't need a new one.

In the dashboard: your balance, how much you have spent, and usage by model and by day.

We run open-weight models on cost-efficient infrastructure and publish fixed USD rates per million input and output tokens for each model — no subscriptions or plans.

The API doesn't recognize the key: it is mistyped, belongs to another service, or was replaced — after Rotate the old key stops working.

Check

The key in your tool is the current one, without spaces, quotes or line breaks, and the header is exactly Authorization: Bearer <key> (Anthropic clients: x-api-key).

Fix

Paste the current key again. Lost it? Press Rotate in the dashboard and update the key everywhere you use it.

Make sure it works

Send a short request: an answer instead of 401 means the key works.

Still failing? Send us

The time of the request, the tool or SDK and the error text — never the key itself.

The request was refused before it reached the model, for one of two reasons. The error text tells them apart: Total limit exceeded — the balance is used up or too low for this request; no active billing account — the request carries no API key.

Check

Your balance in the dashboard, and that the request sends the key as Authorization: Bearer <key> — “Bearer” with a capital B and a space before the key — or as x-api-key. A lowercase bearer or a key without “Bearer” counts as no key.

Fix

Balance used up: press Top up in the dashboard — the same key works again once the top-up is credited. No key: add it to your tool or environment variable, then restart the tool or the terminal.

Make sure it works

In the dashboard the key shows Active, and a short request returns an answer.

Still failing? Send us

If the balance is positive and the key is sent but 402 remains: the time of the request, the model and the full error text.

Requests come faster than allowed. The limit is 120 requests per minute per API key, counted in one-minute windows. A 429 can also mean the service or a model is briefly busy.

Check

The error text: Rate limit exceeded — your key hit the per-minute limit; at capacity, waiting for an available upstream slot or overloaded — the service or the model is busy.

Fix

Wait the number of seconds in the Retry-After header and repeat — the official OpenAI and Anthropic SDKs retry a 429 on their own, twice by default. For agents, limit how many requests run in parallel; if one model stays busy, switch the model field.

Make sure it works

Requests return answers again, and at a steady pace below the limit the 429s stop.

Still failing? Send us

If 429 comes at a low request rate: the time, the model, how many requests you send per minute and the error text.

A 404 means the address is wrong. A 400 that says “not found” means the API doesn't know the model ID.

Check

OpenAI-style clients need the address with /v1: make sure it isn't missing, or doubled because the client adds /v1 itself. The model ID must be complete, with the prefix before the slash, for example zai-org/GLM-5.3-Flash.

Fix

Copy the address and the model ID from the settings for your tool on this page, or take the ID from GET /v1/models.

Make sure it works

A short request returns an answer, and GET /v1/models lists the same ID.

Still failing? Send us

The full address you call, the model ID and the error text.

The model can't take requests right now. The service status at the top of this page and the model cards show the same state.

Check

The service status above: is this model marked as temporarily unavailable?

Fix

Repeat the request in a minute, with growing pauses — a 503 has no Retry-After header; the official SDKs retry it twice on their own. Or switch the model field to another model — the key and the address stay the same.

Make sure it works

The model shows as available again, and a short request returns an answer.

Still failing? Send us

If the status shows the model as available but 503 continues: the time, the model and the error text.

The answer hit a length limit: finish_reason is length. Without max_tokens, a non-streaming answer stops at the model's default length.

Check

The finish_reason field and the max_tokens you send — the docs list the default and the maximum output for each model.

Fix

Set max_tokens for longer answers, up to the model's maximum, or stream the response: without max_tokens, a streamed answer can run up to the maximum.

Make sure it works

The answer ends with finish_reason stop.

Still failing? Send us

If an answer ends with length well below your max_tokens, the stream was interrupted — repeat the request, and if it keeps happening, send us the time, the model and your max_tokens.

Reasoning models think before they answer, so the first token can take a while, and long answers take minutes. The API waits up to 150 seconds for a model to start answering.

Check

Whether you stream the response, and your client's timeout: it should be longer than 150 seconds.

Fix

Stream the response (stream: true): tokens arrive as they are generated, and keep-alive messages hold the connection open. Without streaming, keep the client timeout above 150 seconds — the official OpenAI and Anthropic SDKs wait 10 minutes by default; in Claude Code it is API_TIMEOUT_MS, in Aider --timeout, in OpenClaw timeoutSeconds of the provider. If a streamed answer breaks off, repeat the request: the output of an interrupted answer isn't charged.

Make sure it works

The first tokens appear in the stream, or the full answer arrives within the timeout.

Still failing? Send us

The time, the model, whether you stream, your client timeout and where the answer stopped; the x-request-id response header, if your tool shows it.

Enter your email, and we'll send you a magic link — no password needed. You can also continue with Google, MetaMask, or GitHub.

Each sign-in link works once and for a limited time, and only the newest one opens.

Check

The spam and promotions folders, and the spelling of the address on the login page.

Fix

Request a new link on the login page and open the newest email.

Make sure it works

The link opens the dashboard with your account.

Still failing? Send us

The address you sign in with and when you requested the link — the form on this page works without signing in.

The dashboard shows a key only once. Press Rotate to get a new one — the old key stops working, so update it everywhere you use it.

Share your referral link. You get 10% of every balance top-up made by people who sign up with it — for as long as they use CheapTokens.

In the dashboard, on the Referral program tab: sign up or log in and copy it there.

1.3M+ free tokens at sign-up and 52M+ more after their first top-up of $5 or more.

Each reward is on hold for 7 days after the top-up, then becomes available on your referral balance. Move it to your API balance, or withdraw USDT from $20.