OpenCheap

npx add-mcp 'https://app.opencheap.si/api/v1/mcp'

Connect your agent.Buy every AI model.

Give this to your agent

set up https://opencheap.si/SKILL.md

and let it take it from there.

1Providers

2OpenCheap

3You pay

Anthropic$10.00

Provider A$9.11

Provider B$6.00

Provider C$1.10

Provider D$4.57

Routed to Provider C

claude-sonnet-5Live

List price

$10.00

Our price

$1.10

You save$8.90

Our price and the list price per 1M output tokens, live. The provider routes are illustrative.

840+ models from every major lab. One API key.

• Anthropic
• OpenAI
• Google
• Meta
• DeepSeek
• Mistral
• Qwen
• xAI
• Moonshot

Up to 99% below list on popular models.

Model	List price	Our price	Below list
claude-sonnet-5	$10.00	$1.10	−89%
claude-opus-5.5	$20.00	$12.00	−40%
gpt-5.5	$30.00	$3.90	−87%
gemini-3.5-flash	$9.00	$0.270	−97%
kimi-k3	$15.00	$0.160	−98%
gpt-oss-120b	$0.600	$0.003	−99%
glm-5.3-flash	$0.092	$0.011	−88%
llama-3.3-70b	$0.520	$0.070	−86%
grok-4.7	$6.00	$0.900	−85%
glm-5.2	$6.80	$0.000	−99%

Our price per 1M output tokens, next to OpenRouter's list price. See all 845 prices

Why it's cheaper

Same model. Cheapest provider.

You pick the model. The router finds who serves it for less right now, and you pay our price, shown next to the public list price.

1. kimi-k3price, low → high Provider C1 Provider A2 Provider D3 Provider B4 01

   Every provider, ranked

   The same model is sold by many providers at very different prices. Each request ranks every one that serves it, for your prompt and output size.
2. Provider Cbreaker open Provider Arouted Provider Dstandby 02

   The cheapest healthy one answers

   Your request goes to the cheapest provider that is up. Circuit breakers take out one that fails or slows down until it recovers.
3. Provider A502 error Provider D200 OK Failed before the first byte, so the next one answers. 03

   The next one if it fails

   If a provider errors before the first byte, the next cheapest takes the request, at the same price to you.

Same model id, never a substitutePrices refreshed every five minutesPrompts never stored

How routing works

For agents

Made for agents that buy their own tokens.

Agents call many models, many times a day. OpenCheap lets them connect, pay and stay inside a budget with no person in the loop, at the cheapest price for each call.

Pay per call with x402

no API key · no account

1

POST /v1/chat/completions

{ "model": "claude-sonnet-5", … }

2

402 Payment Required

PAYMENT-REQUIRED: 0.0031 USDC on Base

3

retry with PAYMENT-SIGNATURE

200 OK · x-price-usd: 0.0031

Example amount. The real quote depends on the prompt and the output cap.

• $ set up opencheap.si/SKILL.md

  Set up in one line

  Paste the line into Claude Code, Cursor or Codex. The agent reads it and connects itself.
• Balancelow $1.80$21.80 +20 USDC from its wallet

  Tops up its own balance

  An agent can sign in with its wallet as its own account and add USDC when the balance runs low.
• research-agentthis month $41.20/ $50 Stops at the limit

  Spend caps that hold

  A monthly limit per key, plus daily, weekly or monthly budgets that stop or alert.
• Make a key for my scraper with a $20 monthly limit create_keyDone: key made, $20 a month.

  Run the account over MCP

  Ask in plain words: make a key with a limit, show this week's usage, call a model.
• GET /api/v1/modelsno key claude-sonnet-5$1.10 claude-opus-5.5$12.00 gpt-5.5$3.90 per 1M output tokens

  A catalog it can read

  Every model and its price at one URL, no key needed, so the agent can choose for itself.

Three ways to connect.

FAQ

Questions, answered.

How are your prices set?

Every request is routed to the cheapest healthy source for that model, across several providers. Popular models usually cost less than OpenRouter's list price; the ones that cost more are marked "above list" on the Models page. Prices can change at any time, and the Models page shows OpenRouter's list price next to ours, so you can compare.

Is it really the same model?

Yes. A model id always means that model and version, never a renamed substitute. Like any API, answers can vary from one run to the next, even at temperature 0.

How reliable is it?

Fail-over to the next provider before the first byte when a model has more than one, circuit breakers per provider and model, and a catalog refreshed every five minutes. The latest stress test served 430 of 430 requests at up to 50 concurrent.

Can AI agents pay by themselves?

Yes. Call the API with no key and it answers 402 with the price of that one completion; the agent pays it in USDC from its wallet (x402) and gets its answer in the same exchange. Agents with an API key can also top up their balance from code when it runs low.

Who sees my prompts?

The provider that serves the request processes it under its own terms, and so does any provider it passes the request on to; we store metadata only (model, tokens, latency, cost), never prompt or completion bodies.

Any fees?

No monthly fee and no plan: pay as you go from one balance, by card or in USDC. Credits never expire and cannot be withdrawn.