Kimi K3 on BestRelay: a pay-as-you-go API relay for long inputs

Kimi K3 on BestRelay answers under the name you write

BestRelay is an independent API relay, and its central point is simple: the model you name is the model that answers. A request for Kimi K3 is forwarded to the provider under that name, so nothing is quietly swapped for a cheaper tier on the way through. On a desk where one request carries half a million input tokens, knowing which model produced the reply is the difference between a budget you can read in advance and a bill you can only interpret afterwards.

The line-up is curated rather than swept together: 20 public models, each listed under its own name, each with four prices for input, output, cache write and cache read. Kimi K3 sits in that list at $1.39 per 1,000,000 input tokens, $7.01 per 1,000,000 output tokens and $0.1389 per 1,000,000 cache read tokens. When a provider retires a model id, BestRelay delists it rather than leaving a dead name in the catalogue.

What follows is an example, and it is labelled as one: a sample working day at your own desk, with you at the keyboard and Kimi K3 doing the reading. The point is not to make a case about volume. It is to show what very long inputs add up to once they are measured in tokens rather than impressions, and to show that on BestRelay the arithmetic stays legible from the first request to the last.

You are not signing up for a bundle or a fixed term. You name a model, send a request, and receive an answer from that model, with the charge recorded against the rate card after the call ends. BestRelay is the desk between you and the provider, and the desk is deliberately plain.

A sample day of long inputs, worked through as an example

Here is the example in full. Suppose your working day on Kimi K3 involves 500,000 input tokens and 100,000 output tokens. On BestRelay, that day costs $1.40. Twenty-two such days cost $30.71. If 350,000 of the input is read from cache instead, the day costs $0.96.

These are examples rather than quotes, and each figure answers a different question. The $1.40 figure is the plain case: a heavy day of reading, in which the model is fed a great deal and asked for a compact reply. The $30.71 figure is the same day repeated across a working month, which is where long-input work becomes worth planning. The $0.96 figure is the cached variant, in which much of the material you already supplied is read back rather than processed fresh.

The mechanism behind that last figure is the cache read rate. For Kimi K3, cache read is priced at $0.1389 per 1,000,000 tokens, below the $1.39 input rate. If your day involves a stable body of reference material — a specification, a code base, notes you return to often — much of your input can arrive through the cache, and the day's cost falls accordingly.

Notice what the example does not require. There is no monthly fee, no subscription and no minimum spend attached to it. You work with a prepaid pay-as-you-go balance, and each call is charged once, after it ends, from the tokens the provider reports, at the rate card prices. If the provider rejects a request, the call is recorded at zero; if the provider cannot be reached, nothing is charged.

So the sample day is not a special arrangement. It is the ordinary shape of using BestRelay: name Kimi K3, send the long prompt, and let the usage records show you the model, the tokens and the charge.

Kimi K3's rates beside the rest of the BestRelay catalogue

Long-input work is rarely a single-model habit, which is why the rate card matters more than any one headline figure. Kimi K3 is priced at $1.39 input, $7.01 output and $0.1389 cache read per 1,000,000 tokens. Set that beside GPT-6.1 Sol at $0.2083 input, $1.11 output and $0.0431 cache read, or Claude Sonnet 5.5 at $1.39 input, $7.01 output and $0.0944 cache read.

As an example, the same working day of 500,000 input tokens and 100,000 output tokens costs $0.22 on GPT-6.1 Sol at 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, a difference of $1.78 on the day and $39.26 across the 22 days.

As another example, the same day on Claude Sonnet 5.5 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, a difference of $0.60 on the day and $13.29 across the 22 days.

Those comparisons are examples of arithmetic, not a verdict on which model suits which task. The structural point is firmer: because BestRelay lists each model under its own name with its own four prices, you can put the day you actually plan into the rate card and read the answer before you begin. Other entries include DeepSeek V4 Pro at $1.53 input and $4.58 output, and Gemini 3.8 Flash at $1.25 input and $5.56 output.

Keys, limits, records and the clients Kimi K3 works with

A day of long inputs still runs on ordinary infrastructure. To connect a client to BestRelay you need two settings: the address and a key created under API keys in the dashboard. The details are in the docs, and they are deliberately not printed here, because a key that appears in an article is a key that has stopped being private.

Your account can hold several keys, and each can be voided at any time. One key handles 120 requests a minute, and one account runs 8 calls in flight, which leaves room for a working desk. Every call appears in the usage records with its model, its tokens and its charge, so at the end of the sample day above you can read back what each request cost rather than estimating it.

On privacy, BestRelay keeps the user-written text of requests for 30 days, so abuse can be reviewed and billing questions settled. System prompts and model replies are not stored. That is a short retention list, and it is worth reading before you paste anything sensitive into a long prompt.

Clients have their own rules, and they are sensible ones. Kimi K3 works with tools such as Codex, Cursor, Cherry Studio, Cline and OpenCode, alongside the GPT, DeepSeek, GLM, Qwen and Grok models in the catalogue. Claude models are served only to the official Claude Code and Claude Desktop clients, version 2.1.285 or newer, and Gemini models use only their native protocol, through Gemini CLI or direct generateContent calls. If your habits are already formed, BestRelay meets them where they are.

Balances, top-ups, refunds and referral credit

Each new BestRelay account opens with $0.35 of credit, enough to make a first call and watch the whole loop work: request, response, usage record, charge. As an example, the free credit alone covers about 1.6 million GPT-6.1 Sol input tokens. Top-ups run through Stripe checkout in set amounts of $1.39, $6.94, $13.89, $27.78, $69.44 and $138.89, so you can add what a week needs instead of committing to a figure you have not yet earned.

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. There is no monthly fee, no subscription and no minimum spend, which means the balance belongs to you in the sense that matters: if you stop, you stop paying.

Referrals are handled plainly. 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. As an example, on an invited member's top-up of $13.89, the two commissions are $2.50 and $0.5556.

That is the sample day in full, and it is deliberately unremarkable. You name Kimi K3, you send a long input, you are charged from the tokens the provider reports, and the record shows the model, the tokens and the charge. When you want to put your own day into the same arithmetic, BestRelay is there with a curated line-up, a prepaid pay-as-you-go balance and a $0.35 opening credit.