The vocabulary, without the vocabulary

API key

A long password your software uses to prove a request is yours. Anyone who has it can spend your credit, so treat it like a card number.

Shown once, at creation. We keep only a hash, which means we genuinely cannot recover it for you — a lost key has to be replaced.

Batch

A file of requests run without anyone waiting for them. Useful when you have ten thousand things to summarise and no deadline in the next few seconds.

Bring your own key

Use an account you already have with a provider, and pay them directly. Requests still route, retry and get logged here, but nothing is charged to your Multigrid balance.

Your key is encrypted before it is stored and is never shown again after you enter it.

Cache hit

The identical request was made recently, so the saved answer was returned. It cost nothing and arrived instantly.

Completion

The model's answer. It is charged separately from your question, and usually costs more per word.

Connected app

Another company's app that you have allowed to send requests using your credit, up to a limit you chose. You can cut it off at any time.

Context length

How much the model can read at once, counting your whole conversation and any documents you attach. Go over it and the request is refused rather than truncated.

A 128,000-token context is roughly a 300-page book. Bigger is not automatically better: you pay for every token you put in it.

Credit

Money you have added, which requests are drawn from. It does not expire, and running out stops requests rather than producing a bill.

Embedding

A way of turning text into numbers so a computer can find things that mean the same, not just things spelled the same. It is how search over your own documents is built.

Failover

If a provider is down or refuses the request, we try the next one that serves the same model rather than handing you the error.

Retries against the same provider come first, with a growing delay between attempts. Only then do we move on.

Gateway

One address that forwards your request to whichever provider should handle it, so your code never has to know which one served it.

It also means one bill, one key, and one place that has a record of every request.

Guardrail

A rule that checks requests before they are sent, or answers before they are returned — for example, refusing anything containing a card number.

Latency

How long a request took, end to end.

Ledger

The permanent record of every credit and charge on your account. Nothing in it is ever edited or removed; a correction is added as its own entry.

Your balance is derived from it rather than stored independently, which is what makes it auditable.

Model

One specific AI, like GPT-5.2 or Claude Sonnet 5. They differ in price, speed, and what they are good at — which is why it is worth being able to switch.

Per million tokens

Prices are quoted per million tokens because per-word prices would be a row of zeros. A million tokens is about 750,000 words, or ten novels.

Preset

A saved setup — a model, an instruction, and its settings — that you can call by name. It lets you change which model your app uses without touching the app.

Prompt

Everything you send the model — your question plus any instructions and documents attached to it.

Provider

The company whose computers actually run the model. The same open model is often sold by several of them at different prices and speeds.

Rate limit

The most requests per minute one key is allowed. It exists so a bug in a loop cannot make ten thousand calls before anyone notices.

Rerank

Given a question and a pile of candidate documents, put the useful ones at the top. It is the step that makes search results actually relevant.

Spend limit

A ceiling you set on yourself. Once month-to-date spend would go past it, requests are refused with a clear reason instead of quietly continuing.

Streaming

The answer arrives word by word as it is written, instead of all at once at the end. It costs the same; it just feels far faster.

Telemetry export

Sends a record of each request to a monitoring tool your team already uses, so gateway calls appear alongside the rest of your system.

OpenTelemetry over OTLP/HTTP. Prompt and answer text is never included — only timings, counts and cost.

Throughput

How fast words come out once the answer has started, measured in tokens per second.

Time to first token

How long you wait before the first word appears. It is what makes a model feel quick or sluggish, and it is not the same as how long the whole answer takes.

Only measurable on streamed requests, which is why it is blank on some rows.

Token

The unit models are billed in. A token is roughly ¾ of an English word, so 1,000 tokens is about 750 words — a long email.

Both what you send and what comes back are counted, and they are usually priced differently: output is typically three to five times the price of input.

Vision

The model can look at images and PDFs you attach, not just read text.

Webhook

An address of yours that we send a message to when something happens — for example, when spend crosses a threshold you set.

Still stuck on something? The documentation goes deeper, and the Playground lets you try any of it without writing code.