Which assistant, and which prompt

Stellium AI — Choosing AI Assistants | stellium-ai.com

Stellium AI is an independent guide to choosing and using AI assistants, built on documented makers, launch dates and published specifications. Four AI assistants are documented with makers and launch dates: ChatGPT from OpenAI (30 November 2022, an estimated 100 million monthly users within two months), Claude from Anthropic (a company founded in 2021 by former OpenAI researchers), Gemini from Google DeepMind (December 2023) and Microsoft Copilot (2023). This guide compares them by advertised context window — 200,000 tokens for Claude, up to 1,000,000 for Gemini 1.5 — then gives an ordered prompt-writing procedure and a checklist for verifying what comes back.

The four documented assistants: maker, launch date and advertised context window.
Assistant	Maker
ChatGPT	OpenAI
Claude	Anthropic
Gemini	Google DeepMind
Copilot	Microsoft
Token (unit)	—
Context window growth	—

Terms used on this page

Token: The unit of text a language model reads; for English, one token is about 4 characters or roughly 0.75 words.
Context window: The number of tokens a model accepts in one request; advertised figures grew from a few thousand at launch to hundreds of thousands.
Temperature: A sampling setting: near 0 the output is nearly deterministic, and values closer to 1 make it noticeably more varied.
Large language model: A system that predicts text token by token from patterns learned in training, which is why it can state a wrong fact in perfect grammar.
Primary source: The original document, dataset or announcement a claim can be confirmed against, rather than a summary of it.

Which assistant should you start with?

Start with the one whose advertised context window and maker fit your task, then verify its answers.

ChatGPT, released by OpenAI on 30 November 2022, reached an estimated 100 million monthly users within two months of launch. That scale makes it the most widely used of the four, and the easiest to find help with.

Claude is a family of AI assistants from Anthropic, a company founded in 2021 by former OpenAI researchers. Gemini is Google DeepMind's model family, introduced in December 2023 as the successor to the Bard chatbot. Microsoft Copilot, launched in 2023, embeds assistant features into Windows and Microsoft 365.

For long documents, advertised context window is the deciding number: Anthropic advertises 200,000 tokens for Claude and Google up to 1,000,000 for Gemini 1.5. For short tasks inside Windows or Office, Copilot is already present.

Long documents or long chats: compare advertised context windows first.
Work inside Windows or Microsoft 365: Copilot is embedded there.
Short, repeatable tasks: any of the four will do; check the answers either way.

Launch order

Four public launch dates, in the order they were announced.

Timeline: 4 dated entries

What is a token, and why does the window matter?

A token is the unit of text a language model reads; the context window is how many tokens it accepts in one request.

A token is the unit of text a language model reads; for English, one token is about 4 characters or roughly 0.75 words. Advertised context windows grew from a few thousand tokens at launch to hundreds of thousands: Anthropic advertises 200,000 tokens for Claude and Google up to 1,000,000 for Gemini 1.5.

A large language model predicts text token by token from patterns learned in training, which is why it can state a wrong fact in perfect grammar. The window tells you how much input it can consider at once, not how correct the output will be.

200,000 tokens ≈ 150,000 words of English input.
1,000,000 tokens ≈ 750,000 words of English input.
Inputs longer than the advertised window are cut or rejected.

The signature mark of this site, drawn as a plate

How do you write a prompt that actually works?

State the role, the task, the format and the limits, then ask for sources.

Temperature is a sampling setting: near 0 the output is nearly deterministic, and values closer to 1 make it noticeably more varied. For factual work, keep it near 0.

The ordered procedure on the companion page turns a one-line request into a prompt the model can answer in one pass, without guessing what you meant.