General assistant
Your default for writing, thinking, summarising and everyday questions.
What decides it: Prose quality, how well it follows instructions, and whether it admits uncertainty instead of inventing.
- ChatGPT
- Claude
- Gemini
Course reference
Every fact that goes stale lives on this page, and nowhere else in the course.
Last verified: 13 September 2026. Model names, prices and rankings change every few months. The lessons teach you how to choose; this page tracks what is currently worth choosing. If the date above is more than a quarter old, trust the live sources at the bottom over the names here.
Your default for writing, thinking, summarising and everyday questions.
What decides it: Prose quality, how well it follows instructions, and whether it admits uncertainty instead of inventing.
A contract, a thesis, a year of meeting notes — anything you cannot paste in pieces.
What decides it: Context window size, and whether it quotes the source rather than paraphrasing it.
The answer depends on something recent, and you need to check where it came from.
What decides it: Live web access and real, clickable citations you can verify.
Anything going out under your name.
What decides it: How well it holds a voice profile across a long piece without drifting to house style.
Numbers matter and a wrong one is expensive.
What decides it: Whether it EXECUTES code rather than predicting the answer. Always ask to see the working.
Illustration, mockups, thumbnails, concept art.
What decides it: Prompt control, text rendering inside images, and licensing terms for commercial use.
Narration, dubbing, cloning your own voice for scale.
What decides it: Naturalness, language coverage, and consent controls around cloning.
Short clips, talking-head presenters, B-roll.
What decides it: Clip length, shot consistency, and whether the output survives being watched twice.
A repeated sequence across apps that should run without you.
What decides it: App coverage, pricing model (per task vs per operation), and whether you can self-host if privacy demands it.
Multi-step work where the AI decides the next step, not you.
What decides it: Tool-calling reliability, memory across runs, and MCP support for connecting your systems.
The data genuinely cannot leave your machine.
What decides it: What your hardware can run, and how much capability you are willing to trade for privacy.
A written list is out of date the week it is published. These are updated continuously — use them before any decision that costs money.
Tools churn; the questions do not. Whatever is winning this quarter, you still choose by asking: does it need current information, how long is the input, does a human read the output as mine, does a wrong number cost me, and can this data leave my machine? Those five questions have survived every model release so far.
The decision tree