Pricing pages and capability claims in this category use a compact vocabulary, and several of the words carry more precision than they appear to. These definitions are what make the rest readable.
The unit of text a model processes and the unit you are billed in. Roughly four characters of English, so about three quarters of a word, though this varies by language and content.
Reading it well: different models tokenise the same text differently, so identical rates can produce different bills.
What you send and what the model generates, billed at separate rates. Output is consistently the more expensive side, commonly three to six times the input rate.
Reading it well: a headline input price says little if your workload generates long answers.
Reusing previously processed prompt content across requests at a reduced rate. Anthropic documents cache reads at 10% of the base input price, with a write premium.
Reading it well: only pays off with genuine repetition. Mechanics and discounts differ by provider.
Submitting work for asynchronous processing at a discount, commonly around 50%, in exchange for giving up an immediate response.
Reading it well: the single easiest saving available for anything that is not interactive.
A higher rate applied above a token threshold on some models. Others charge flat across the whole window.
Reading it well: this can outweigh the headline rate difference for document-heavy workloads.
The quota band your account sits in, which rises with usage history or by arrangement. Separate from price.
Reading it well: a new account will not have production throughput on day one, whatever it is willing to pay.
The maximum tokens a model can consider at once, covering your prompt, any documents and the conversation so far.
Reading it well: a ceiling rather than an allowance. You pay for what you send, not for the size of the window.
Handling more than text. The word covers input, output or both, and providers use it loosely.
Reading it well: ask which specific modes go in and which come out, since the label alone does not say.
Allowing a model to work through a problem at length before answering. Usually improves difficult tasks and consumes more tokens and time.
Reading it well: those extra tokens are usually billed, so this is a cost decision as much as a quality one.
The point after which a model has no training data. It does not always know what it does not know.
Reading it well: search or retrieval closes the gap. Without one, treat recent-event answers with suspicion.
A confident, fluent, incorrect output. Present in every current model to some degree.
Reading it well: a property of the technology, not a defect specific to any product. Plan verification accordingly.
The component splitting text into tokens. Changes between model generations.
Reading it well: Anthropic documents its newer tokenizer producing roughly 30% more tokens for the same text, which changes costs at an unchanged rate.
A model requesting that your code run something and returning the result into its reasoning. The basis of most agent behaviour.
Reading it well: tool definitions consume input tokens on every request, which adds up at volume.
A loop where a model plans, calls tools, observes results and continues with limited supervision. There is no agreed technical definition.
Reading it well: ask what it is permitted to do without asking. That is the meaningful specification.
Fetching relevant material and supplying it in the prompt so answers rest on your documents rather than training data.
Reading it well: usually cheaper and more checkable than pushing everything into a long context window.
Further training a model on your own examples to adapt its behaviour. Availability differs by provider and is standard for open weights.
Reading it well: often unnecessary. Prompting and retrieval solve a large share of what people reach for fine-tuning to fix.
Constraining responses to a defined schema so software can parse them safely.
Reading it well: check whether the schema is enforced or merely requested. It changes how much validation you write.
Model parameters published for download. Not the same as open source, since the licence may restrict use.
Reading it well: read the specific licence. Several widely used open-weight models carry conditions on commercial use.
Leading on some benchmark at some moment, usually as measured by whoever is making the claim.
Reading it well: ask which benchmark, whose measurement and what date. Without all three it is not checkable.
Informal shorthand for the most capable current models from major labs. No defined threshold.
Reading it well: a marketing category rather than a specification.
Almost always a benchmark result, and frequently one where the harness mattered as much as the model.
Reading it well: your codebase and your tooling decide this, not a leaderboard.
Usually means administrative controls, certifications and a support agreement exist. Not a defined standard.
Reading it well: ask which certifications specifically, and whether the terms permit your use.