Glossary

Agent
A model given tools and a goal, left to loop — deciding what to do next and doing it — until the job is done or it gives up.
Full entry
Agent loop
Plan, call a tool, read the result, decide again. The cycle an agent repeats until it finishes or is stopped.
Full entry
AI21
An Israeli lab behind the Jamba models, built for very long inputs.
Full entry
Alibaba
Publishes the Qwen models, a large open-weight family covering text, images and code.
Full entry
Amazon
Builds the Nova models and resells most other vendors’ through Bedrock, its managed model service.
Full entry
Anthropic
The company behind the Claude models. Publishes the most complete pricing of any vendor read here, including what writing to the cache.
Full entry
ByteDance
TikTok’s parent company, which publishes the Doubao and Seed models.
Full entry
Cache write
The charge for putting a prefix into the prompt cache. Usually more than the input rate, not less.
Full entry
Chain of thought
Asking a model to work through its reasoning before answering, which measurably improves accuracy on multi-step problems.
Full entry
ChatGPT
OpenAI’s chat assistant, on the web, desktop and phone. The usual first encounter with a large language model.
Full entry
Claude
Anthropic’s assistant, on the web, desktop and phone.
Full entry
Claude Code
Anthropic’s coding agent, run in a terminal or inside an editor. It reads and edits a whole repository rather than one file.
Full entry
Codex
OpenAI’s coding agent, run from the terminal or delegated to in the cloud.
Full entry
Cohere
A Canadian company selling models aimed at business use, particularly search and retrieval over a company’s own documents.
Full entry
Computer use
A model driving a screen directly — looking at pixels, moving a cursor, typing — instead of calling an API.
Full entry
Context rot
The tendency of a model to attend less to material buried in the middle of a long context than to its start or end.
Full entry
Context window
The total tokens a model can consider at once: your prompt, the conversation so far, and the answer it is writing.
Full entry
Copilot
Microsoft’s assistant brand. GitHub Copilot writes code in an editor; Microsoft 365 Copilot sits inside Word, Excel, Outlook and Teams.
Full entry
DeepSeek
A Chinese lab whose models are notably cheap per token and published as open weights, which is why they turn up in so many cost comparisons.
Full entry
Embedding
A list of numbers representing a piece of text, arranged so that similar meanings land near each other.
Full entry
Evals
A repeatable test set you score a model against, so a change can be shown to help rather than assumed to.
Full entry
Few-shot prompting
Showing the model two or three worked examples of what you want, instead of describing it.
Full entry
Fine-tuning
Continuing training on a smaller, curated set of examples to specialise a model or teach it a format.
Full entry
Frontier model
The largest and most capable model a vendor currently offers, priced accordingly.
Full entry
Gemini
Google’s assistant, on the web and phone and woven through Workspace — Docs, Gmail and Sheets.
Full entry
Google
The company behind the Gemini models, sold through Google AI Studio and Vertex AI.
Full entry
Grok
xAI’s assistant, reached on its own site and inside X.
Full entry
Hallucination
A fluent, confident, entirely invented answer. The failure mode that matters, because it looks exactly like a correct one.
Full entry
Inference
Running a trained model to get an answer. Everything you pay for by the token is inference, not training.
Full entry
Knowledge cutoff
The date after which a model saw no training data. It knows nothing later unless you tell it or it can search.
Full entry
Large language model
A model trained to predict the next token over enormous amounts of text, which turns out to be enough to do a great deal.
Full entry
Le Chat
Mistral’s assistant, on the web and phone.
Full entry
Llama
Meta’s family of open-weight models. Free to download and run on your own machine; several hosting companies also sell access by the token.
Full entry
Meta
The company behind the Llama models, which it publishes as open weights you can download and run yourself.
Full entry
Microsoft
Sells other companies’ models through Azure alongside its own smaller Phi models, and puts them behind the Copilot name across Windows and.
Full entry
MiniMax
A Chinese lab whose models cover text, speech and video.
Full entry
Mistral AI
A French lab publishing both open-weight and paid models, and the usual answer when a project has to stay inside Europe.
Full entry
Model Context Protocol
An open standard for connecting models to tools and data, so one integration works across many applications.
Full entry
Moonshot AI
A Chinese lab behind the Kimi models, known for very large context windows.
Full entry
Multimodal
A model that accepts more than text — images, audio or video — in the same prompt.
Full entry
Open weights
A model whose trained parameters are published, so anyone can run it on their own hardware.
Full entry
OpenAI
The company behind the GPT models and ChatGPT. Sells access through its own API and through Microsoft Azure.
Full entry
Per million tokens
How model prices are quoted: dollars per million tokens, charged separately for what you send and what comes back.
Full entry
Pretraining
The first and most expensive stage: learning to predict the next token across a very large body of text.
Full entry
Prompt
Everything you send the model on one call: instructions, context, history and question together.
Full entry
Prompt caching
Paying a reduced rate for a prefix the vendor has already processed, instead of full price for sending it again.
Full entry
Rate limit
A cap on how much you may send in a window, usually counted in requests and tokens per minute.
Full entry
Reasoning model
A model trained to produce a long internal working-out before its answer. More accurate on hard problems, slower and dearer.
Full entry
Reinforcement learning from human feedback
Training a model against human judgements of which answer is better, rather than against a single correct answer.
Full entry
Retrieval-augmented generation
Looking up relevant documents and putting them in the prompt, so the model reads facts rather than recalling them.
Full entry
System prompt
Standing instructions sent ahead of every conversation, setting the role, rules and format.
Full entry
Temperature
How much randomness to allow when picking each token. Low is repeatable and dull; high is varied and unreliable.
Full entry
Token
The unit a model reads and writes. Roughly three quarters of an English word, so 1,000 tokens is about 750 words.
Full entry
Tokenizer
The fixed table that chops text into tokens. Decided before training and never changed afterwards.
Full entry
Tool use
Letting a model call functions you define — search, read a file, send a request — and read back what they return.
Full entry
Vibe coding
Building software by describing what you want and iterating on what the model produces, rather than writing it line by line.
Full entry
xAI
The company behind the Grok models.
Full entry
Zhipu AI
A Chinese lab behind the GLM models.
Full entry

Calculator

What will this actually cost?

Most pricing pages multiply tokens in by a rate and tokens out by another. On anything agentic that is the small half of the bill. Set a workload below and this prices the prompt cache too — reads and writes.

Your workload

Input tokens per call

Output tokens per call

Calls per day

Prompt cache

Cache hit rate: 90%

Cache writes per day

A coding agent re-sends its whole context on every turn. The cache makes that cheap to read — and each write still costs, at up to twice the input rate.

Cost per day for the workload above. A model that does not publish a figure this workload needs shows what is missing instead of a number.
Model	Vendor	Per day	Per call	Where it goes

Why the cache write is the line that matters

A prompt cache lets a vendor skip re-processing a prefix you have sent before, and charges you much less for those tokens — often a tenth of the input rate. The part that is easy to miss is that putting something into the cache costs more than sending it normally: vendors that publish the figure charge around 1.25× the input rate for a five-minute entry and 2× for an hour.

So a workload that rewrites its cache often can cost more than one that never caches at all. That is not an edge case — it is what happens when an agent edits its own context between turns. Set the cache writes above to a high number and watch the total move.