BestRelay API proxy pricing for DeepSeek V4 Pro usage

BestRelay forwards each request to the model name the caller wrote

A side project rarely arrives with a tidy, generous budget. BestRelay suits that kind of workload because it is an independent API relay with one plain rule: it forwards each request to the model provider under the model name the caller wrote, so the model named is the model that answers and nothing is quietly swapped for a cheaper tier.

The line-up is curated rather than sprawling. BestRelay lists 20 public models, each under its own name with four prices — input, output, cache write and cache read — so the cost of a call can be read before it is made. When a provider retires a model id, BestRelay delists it, which keeps the catalogue free of dead entries and keeps a balance from being spent on calls that cannot land.

That curation is the quiet reason a small project can treat a model choice as a decision rather than a gamble. If the sheet says Claude Sonnet 5.5, that is the model that receives the request; if it says GPT-6.1 Sol, that is the model that replies. There is no hidden substitution, no cheaper tier beneath the name, and no separate sales pitch for a bundle of models nobody asked for.

For a builder, this matters twice over. At the outset, it removes the unpleasant discovery that a familiar label was routed somewhere else. Later on, it makes the rate card a usable budget: every figure is per 1,000,000 tokens, and every figure is attached to a model you chose yourself. BestRelay does not ask you to trust a marketing page; it asks you to read a table. That is a modest promise, but a modest promise is one a side project can actually plan around.

A prepaid balance charges only the tokens a finished call used

BestRelay works on a prepaid pay-as-you-go balance. There is no monthly fee, no subscription and no minimum spend, which suits work that begins small and may not continue at all. You put money in, the balance stands ready, and spending follows usage rather than the calendar.

The accounting is deliberately plain. Each call is charged once, after it ends, from the tokens the provider reports, at the rate card prices for the model that ran. Nothing is held in advance, so an unfinished idea does not sit behind a standing charge. The balance moves because tokens moved, and the amount that leaves it can be traced to a specific call.

Errors are treated with the same plainness. Requests the provider rejects are recorded at zero, and when the provider cannot be reached, nothing is charged. A failed attempt therefore costs nothing beyond the attempt itself, which is a reasonable position for a project that runs on small, repeated experiments rather than on carefully rehearsed batches.

The dashboard keeps the receipts. Every call appears in the usage records with its model, tokens and charge. An account can hold several keys, and any key can be voided at any time, so a key issued for one job can be set aside without disturbing the others. Limits are stated plainly too: 120 requests a minute on one key, and 8 calls in flight on one account. BestRelay is not asking a small project to buy scale it has no use for, and the records make the spending legible without any further effort.

Top-ups in small Stripe amounts keep the balance close to the work

Adding credit is done through Stripe checkout, and the amounts suit modest budgets: $1.39, $6.94, $13.89, $27.78, $69.44 and $138.89. A short experiment can begin at the smaller end of that range, while a project with steady traffic can step up without arranging anything unusual. BestRelay does not sell a period pass; it holds a balance you fill when you decide to, and nothing renews on your behalf.

The refund position is equally direct. Unused credit can be refunded on request within 13 days of the top-up, less what was consumed and any payment-processor fees. Gift credit and commission cannot be refunded. For example, a $13.89 top-up left largely untouched is still recoverable within that window, once the consumption and the processor's fee are set aside.

There is also a referral scheme, which is worth understanding because it changes the arithmetic of a top-up for the person who invited you. You receive 18% of the top-ups of people who joined through your link, and 4% of the top-ups of the people they invited. Commission can be moved into the balance, or withdrawn to Alipay from $20.83. For example, on an invited member's top-up of $13.89, the two commissions are $2.50 and $0.5556.

None of this requires a decision in advance. The smallest top-up is small enough to be treated as a test, and the usage records will say what it bought. BestRelay leaves the choice with the person paying, which is the arrangement a careful side project would ask for anyway.

Claude Sonnet 5.5 and GPT-6.1 Sol show what the rate card is worth

Two examples from the catalogue make the point. For Claude Sonnet 5.5, as an example, a day of 500,000 input and 100,000 output tokens costs $1.40 on BestRelay, and 22 such days cost $30.71; with 350,000 of the input read from cache, the day costs $0.94. At the official list price, the same day is $2.00, the 22 days $44.00 and the cached day $1.33. The difference is $0.60 on the day and $13.29 over the 22 days.

For GPT-6.1 Sol, as an example, a day of 500,000 input and 100,000 output tokens costs $0.22 on BestRelay, and 22 such days cost $4.74; with 350,000 of the input read from cache, the day costs $0.16. At the official list price, the same day is $2.00, the 22 days $44.00 and the cached day $1.33. The difference on the day is $1.78 and on the 22 days $39.26.

The unit prices behind those examples are modest: $1.39 for Claude Sonnet 5.5 input and $7.01 for its output, $0.2083 for GPT-6.1 Sol input and $1.11 for its output, per 1,000,000 tokens. A new account opens with $0.35 of credit, which, for example, covers about 1.6 million GPT-6.1 Sol input tokens. A side project can therefore begin with the free credit alone and consider a top-up only once the usage records show something worth continuing. Caching, meanwhile, is priced on its own line, so repeated context can be billed as cache reads rather than as fresh input.

Client fit, keys and privacy decide what a relay can honestly promise

A relay is only useful if it fits the tool already in use. Claude models are served only to the official Claude Code and Claude Desktop clients, version 2.1.285 or newer; they do not work in Cursor, Cherry Studio, Cline, OpenCode, any SDK or cURL. Those tools, like Codex, pair only with GPT, DeepSeek, Kimi, GLM, Qwen and Grok.

Gemini models use only their native protocol, through Gemini CLI or direct generateContent calls, and are not OpenAI-compatible. Connecting takes two settings: the address and a key created under API keys in the dashboard; the details are in the docs. BestRelay keeps this narrow on purpose, because a promise that stretches across incompatible tools is a promise that eventually breaks.

On data, the position is stated in plain terms. The user-written text of requests is kept for 30 days to review abuse and settle billing questions; system prompts and model replies are not stored. BestRelay asks for no more than it needs to run the account and settle the charges, and the usage records show the model, tokens and charge for every call that counted.

Taken together, the case for a small top-up is unglamorous and fairly convincing: the model you name answers, the balance is prepaid, the charge follows reported tokens, and the receipt sits in the usage records. If that suits the way you work, BestRelay is worth trying with the credit already in your account.