Field guide · Updated October 2026
The AI model field guide, without the hype
There are dozens of models and new ones arrive monthly. Instead of memorizing brand names, teach students to ask: what kind of model is this, what is it good at, what does it get wrong, and how do I check?
Why not a ranking?
The “best” model changes every few weeks.
Leaderboards can shift by a few points between releases, and different scoreboards rank models differently. For the live picture, use our AI race tracker, which updates automatically. This guide covers what stays true.
Nine types
Think in kinds of models
These labels overlap: they mix what a model does, what it works with, how big it is, how it is shared and how it behaves. An open-weight model can also be small, multimodal and part of an agent, and most real products combine several.
Type 1
Chat models (LLMs)
How it works: Read huge amounts of text, then predict the next word, over and over. Tuned with human feedback to be helpful.
Good at: Explaining, summarizing, rewriting, brainstorming, drafting questions.
Weak at: Facts it never saw or half-remembers, invented sources, flattering agreement, exact counting.
Try in class: Ask for three explanations of photosynthesis for a ten-year-old. Students rank them.
Type 2
Reasoning (“thinking”) models
How it works: Spend extra computation writing step-by-step work before answering. Trained with rewards on problems whose answers can be checked, like math and code.
Good at: Multi-step math, logic puzzles, planning, code.
Weak at: Slower and costlier. Can still be confidently wrong, and the visible “thinking” is not always a faithful record of how it got there.
Try in class: Give the same word problem to a fast mode and a thinking mode. Compare answers and time.
Type 3
Multimodal models
How it works: Take in more than text: images, audio, sometimes video.
Good at: Describing a photo, reading a chart, transcribing speech.
Weak at: Missing small details, misreading graphs, confident guesses about what it cannot see.
Try in class: Show a graph. Ask for a description. Students list what it missed.
Type 4
Image generators
How it works: Many start from random noise and refine it toward a picture that matches the prompt (diffusion). Others build the image piece by piece, and some products combine methods.
Good at: Concept art, illustration ideas, style exploration.
Weak at: Text inside images, counting objects, hands, bias in who appears for words like “scientist.”
Try in class: Prompt “a scientist” ten times. Tally who appears. Discuss why.
Type 5
Video generators
How it works: Apply a similar process across time.
Good at: Short clips and visual ideas.
Weak at: Physics that does not hold up, people and objects that change between frames.
Try in class: The teacher prepares five-second clips. Students spot the impossible physics.
Type 6
Music and voice models
How it works: Generate audio from text, or imitate a voice.
Good at: Jingles, sound effects, accessibility tools.
Weak at: Copyright and consent questions. Voice cloning enables fraud.
Try in class: Make a jingle for a class rule. Discuss who owns it. Never clone a real student or teacher’s voice.
Type 7
Agents and computer-use
How it works: A language model in a loop with tools: it clicks, browses, runs code, then checks the result.
Good at: Multi-step tasks like research or booking, under supervision.
Weak at: Errors that pile up, and hidden instructions in web pages (prompt injection).
Try in class: Teacher demo only: an agent plans a fictional field trip. The class audits each step.
Type 8
Small and open-weight models
How it works: Models whose files you can download. Smaller ones run on a laptop, even offline.
Good at: Privacy-sensitive work, experiments, learning how models work.
Weak at: Less capable than the biggest closed models. Running them well takes some skill.
Try in class: Run a small model on a school laptop. Compare it with a big one on the same prompts.
Type 9
World models
How it works: Research systems that predict how a scene changes so they can simulate environments.
Good at: Mostly demonstrations and robotics research.
Weak at: Early-stage. Not a classroom tool today.
Try in class: Watch a demo. Ask: does it understand physics, or imitate video?
The major labs
Who makes what, and who may use it
Open-weight and closed models are different trade-offs: open ones can run privately, closed ones are usually more capable.
| Lab | Model family | Weights | Student age rules | Education features |
|---|---|---|---|---|
| Anthropic | Claude | Closed | Consumer accounts are 18+. Under-18 use is through other products built on it. | Learning mode (Socratic style) |
| OpenAI | ChatGPT / GPT | Closed (some open-weight releases) | 13+ with parental consent for ages 13–17 | Study Mode, parental controls, ChatGPT Edu |
| Gemini, Gemma | Closed (Gemma is open-weight) | Varies by country, account and feature. Consumer accounts are generally 13+ (older in some countries); parents can turn on supervised access for younger children in some regions (not the EEA, Switzerland or UK). School accounts follow the administrator’s settings. | Guided Learning, NotebookLM, Gemini for Education | |
| Meta | Llama and successors | Llama was open-weight. Check the license of newer models. | Varies by app | None specific |
| xAI | Grok | Current hosted models closed (Grok-1 weights released in 2024) | Check current terms. Not recommended as a classroom default. | None specific |
| DeepSeek | DeepSeek | Open-weight | Hosted app raises data-privacy questions. For schools, prefer running it locally or avoid it. | None specific |
| Alibaba | Qwen | Many open-weight models | Varies | None specific |
| Moonshot | Kimi | Open-weight models | Varies | None specific |
| Mistral | Mistral | Some open-weight | Varies | None specific |
Snapshot from October 2026, compiled from company help pages and news reports. Product names, ages and features change often, so check each company’s current terms before any student uses a tool. “Open-weight” means the model files can be downloaded, which is not the same as fully open source.
Strengths and limits
The “jagged frontier” in plain words
AI can be brilliant at one task and fail at a similar-looking one. You usually cannot tell which from the outside.
Usually good at
- Explaining a concept several ways
- Summarizing a document you give it
- Drafting, rewriting and translating
- Writing and explaining code
- Generating practice questions
Still tricky
- Made-up facts and sources: realistic but invented citations, quotes and names
- Agreeing too much: researchers call it sycophancy
- Exact counting and decimals: models see chunks of text, not letters
- Riddles that look like famous ones: they match the pattern, not the new details
- Recent events: a model only knows up to its training cutoff unless it searches
- Hidden instructions: text in a web page or document can steer an AI that reads it
- Bias: in who appears in images and which examples it picks
The evidence. In a Harvard/BCG study of 758 consultants, people using AI finished more tasks, faster, with higher quality on tasks inside AI’s abilities, but were 19 percentage points less likely to be correct on a task just outside them (Dell’Acqua et al., 2023). Not knowing where the edge is, is the problem. Learn to find it →
Interactive
Which tool for which job?
Pick a task. See the best kind of model, and the catch.
Open vs. closed
What does “open” mean?
Closed
Use it through the company
You send your text to the company’s servers. Usually the most capable, with the least control and the least visibility into how it works.
Open-weight
Download the model
You can run it yourself, even offline, and study it. Licenses differ: some allow almost any use, others have limits.
Fully open source
Weights, code and detailed data information
Rare. The Open Source Initiative’s definition requires the weights and code, detailed information about the training data (the full dataset may not be shareable) and freedom to use, study, modify and share. Most “open” models share only the weights.