Claude Opus 5 · relay online

Buy tokens.
Use them your way.

Buy 1M–5M token packs for token-metered API access. Your OpenAI-compatible key is issued after payment clears. No subscription.

1M–5M token packs
Key lands in seconds
Hard token quota per key
Works with any OpenAI SDK

Grab a pack

tokens

Unfinished order — on this device.

How many million tokens?$3.50 per 1M

TOTAL $3.50 × 1M tokens

$3.50

I agree to the Terms, authorize this payment, request immediate delivery, and acknowledge the refund/dispute policy.

Tick the box to unlock checkout.

Your key is live!

Save this now — it is shown once. After you leave, it can only be recovered from this browser session.

API KEY

—

BASE URL

—

TRY IT RIGHT NOW

Three steps. No onboarding.

No account, no email confirmation, no provisioning queue. Payment clears, key exists.

1
Pick a pack

Choose a 1M–5M token pack. The total is the selected millions multiplied by the displayed per-1M price.
2
Pay with PayPal

PayPal handles the entire payment leg. Card numbers never touch this service — we only ever see a capture ID.
3
Get your key

It appears right on this page with your quota attached. Point any OpenAI-compatible client at the base URL and go.

Drop-in for anything that speaks OpenAI.

Swap the base URL and the key. Chat completions behave exactly how your code already expects.

curl https://tokenrelay-staging.openplss11.workers.dev/v1/chat/completions \
  -H "Authorization: Bearer $TOKENRELAY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "messages": [{ "role": "user", "content": "Say hi in five words." }]
  }'

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["TOKENRELAY_KEY"],
    base_url="https://tokenrelay-staging.openplss11.workers.dev/v1",
)

resp = client.chat.completions.create(
    model="claude-opus-5",
    messages=[{"role": "user", "content": "Say hi in five words."}],
)

print(resp.choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.TOKENRELAY_KEY,
  baseURL: "https://tokenrelay-staging.openplss11.workers.dev/v1"
});

const res = await client.chat.completions.create({
  model: "claude-opus-5",
  messages: [{ role: "user", content: "Say hi in five words." }]
});

console.log(res.choices[0].message.content);

Sandbox. Chat right here.

Paste your key and talk to Claude Opus 5 without writing a line of glue code. Streams as it types, formats code the instant a fence opens, and your key is kept in this tab.

Connection

API KEY

BASE URL

MODEL

Your key is kept in this tab only and is sent directly to TokenRelay as Authorization: Bearer. It is not stored by the storefront.

System instructions

PERSONA / RULES

Saved on this device and sent as the first system message of every turn.

Tuning

TEMPERATURE 0.7

Stream tokens as they arrive

Opus, on tap claude-opus-5 · no reasoning

Idle

Enter sends · Shift+Enter adds a line · token-metered key loaded

Good questions.

How are token credits counted?

Token mode starts from usage reported by the upstream provider for successful responses, but we do not make you pay for prompt text you did not supply. TokenRelay credits up to 6,000 prompt tokens for provider-injected overhead plus an additional fixed 200 prompt tokens for relay-added compatibility/context normalization on every successful request, then bills the remaining prompt tokens plus all completion/output tokens. The combined prompt credit is capped at 6,200 tokens and never makes usage negative. For streamed responses, TokenRelay explicitly requests a final usage report so the temporary reservation can be reconciled to billable usage.

Can the key go over the token pack I bought?

No intentional overage. TokenRelay keeps a hard quota on the key and temporarily reserves enough quota for an in-flight request. If the remaining balance cannot cover that reservation, the request is rejected before it is sent upstream. If final provider usage is unavailable or a stream ends ambiguously, the reservation can be charged as a fail-closed accounting measure.

Do token credits expire?

They are not time-metered. The key keeps working until its purchased token quota is consumed, or until access is suspended or revoked for abuse, fraud, a refund, reversal, dispute, or chargeback.

Where do I put the key?

Send it as Authorization: Bearer <key> to https://tokenrelay-staging.openplss11.workers.dev/v1. Any OpenAI-compatible SDK, agent framework, or editor plugin that lets you override the base URL can use the active relay.

I paid but never saw a key.

Do not pay again. This page keeps a one-time delivery token for your order in the browser session. Reload and use Recover key, or use Retry delivery in the status area. If delivery still fails, contact support with the PayPal order ID.

What do you store?

The order record includes the token-pack size, amount, PayPal capture ID, payer identifiers returned by PayPal, delivery metadata, and the issued key encrypted at rest. No card data is sent to or stored by ModelRelay.

Terms of purchase

Version 2026-08-25-v26.14 — the whole thing, in plain language.

Service modes and the active offer. ModelRelay supports two separate products. Call mode provides per-successful-response API call credits through ModelRelay. Token mode provides a purchased token quota through TokenRelay. The storefront only sells the mode shown as the active offer at checkout; the order records that mode and fulfillment must return the matching API-key family. Prices, quantities, and the API base URL shown at checkout are part of the purchase description.

Call-mode metering. In call mode, one successful upstream model response consumes one call credit. Requests that fail at the upstream/provider layer and are returned as provider errors are not intended to consume a call credit. Call mode is not sold or billed by token count. Model availability, context limits, output limits, latency, and provider behaviour may change.

Token-mode metering. In token mode, the API key has a hard purchased token quota. For each successful request, the relay starts from token usage reported by the upstream provider and applies two non-billable prompt/input credits: up to 6,000 tokens for provider-injected overhead and a fixed 200-token credit for relay-added compatibility/context-normalization instructions. These credits do not reduce completion/output usage and together cannot reduce billable prompt/input usage below zero. Any reported prompt/input tokens above the combined 6,200-token credit remain billable, and all completion/output tokens remain billable. The relay temporarily reserves a conservative billable-token budget while each request is in flight so concurrent requests cannot overspend the key. For streamed responses, final usage reporting is explicitly requested. A request may be rejected before execution if the remaining quota cannot cover the reservation. If provider usage is missing, a stream terminates ambiguously, or a relay process fails after upstream work may have occurred, some or all of the billable reservation may be consumed as a fail-closed accounting measure. Token mode does not intentionally permit usage beyond the purchased quota.

Third-party AI service. ModelRelay provides managed access to third-party AI infrastructure and is not affiliated with the underlying model or upstream providers unless expressly stated. Provider/model availability, rate limits, token accounting, context limits, output limits, latency, safety behaviour, and model behaviour may change or be interrupted outside ModelRelay's control.

Payment authorization. By completing a live checkout, you represent that you are authorized to use the PayPal account and/or payment method used for the transaction and authorize the displayed charge in USD. PayPal Sandbox/test checkout is not a live payment and may behave differently from production.

Immediate digital delivery. For a live purchase, you request immediate performance and delivery of the digital service after successful payment. Delivery is considered completed when the API credential for the purchased mode is issued and made available through the checkout or recovery flow. Use of that credential is additional evidence that the service was received.

Refund policy. Except where required by applicable law or payment-provider rules, purchases of immediately delivered digital API access are final and non-refundable after delivery. Consumed call credits and consumed token quota are not refundable. Nothing in these Terms limits rights that cannot legally be waived.

Disputes, refunds, reversals, and chargebacks. Where permitted, you agree to contact ModelRelay first about delivery, access, billing, or service issues before initiating a payment dispute or chargeback. If a refund, reversal, dispute, or chargeback is initiated, the associated API credential may be immediately suspended or revoked while the matter is investigated. ModelRelay may contest claims that conflict with transaction, delivery, or usage records and may provide relevant records to PayPal, payment processors, card networks, financial institutions, or other parties involved in resolving the claim.

Transaction, cooldown, and usage records. For fraud prevention, billing verification, support, dispute resolution, abuse prevention, and purchase cooldown enforcement where enabled, ModelRelay may retain transaction/service-access records including payment transaction identifiers, delivery timestamps, API credential identifiers or hashes, IP addresses, country, user-agent information, request timestamps, model identifiers, response status, token usage or call-credit consumption, and pseudonymous HMAC hashes derived from the buyer cookie and IP. Live checkout may enforce a post-purchase cooldown; test/Sandbox environments may temporarily disable that control. A first-party Secure, HttpOnly, SameSite cookie may be used for cooldown and anti-abuse purposes and is not an advertising or cross-site tracking cookie. Raw API keys are not stored in request logs.

API credential responsibility. Your API key is a bearer credential. You are responsible for keeping it secret and for usage made with it. Sharing, resale, public exposure, abuse, fraud, or attempts to bypass service limits may result in suspension or revocation.