About this series
This is a free, nine-part guide to using AI well, written for people who are not technical. You do not need to code. You only need to be curious and willing to try things.
The series works with any AI tool: ChatGPT, Claude, Gemini, Grok, DeepSeek, or whatever comes next. The tools change every few months. The habits in these guides do not.
| Part | What you will learn |
|---|---|
| 0 | What AI tools actually are, who makes them, and what they cost |
| 1 | Using AI for research and making sense of data |
| 2 | Using AI for everyday work: emails, reports, presentations |
| 3 | Where AI goes wrong, and how to catch it |
| 4 | Building your own custom assistant |
| 5 | A step-by-step build you can copy |
| 6 | Prompting, retrieval, or fine-tuning: which one you need |
| 7 | Building an AI that sounds like you |
| 8 | Pointing AI at something you care about |
Read Part 0 first. The rest can be read in any order.
Models vs. harnesses
Every AI product has two parts: a model, which does the thinking, and a harness, which is everything built around it so you can use it.
Think of a restaurant. The model is the chef. The harness is the restaurant: the menu, the dining room, the waiters, the recipes the chef is told to follow. You never talk to the chef directly. You order from the menu, and the restaurant decides how your order reaches the kitchen.
| Model | Harness | |
|---|---|---|
| What it is | The AI "brain," trained on huge amounts of data | The app, website, or tool wrapped around a model |
| What it does | Learns from data and produces answers | Decides what to ask the model, in what order, and what to do with the answer |
| Who makes it | A handful of large labs | Thousands of companies |
| Examples | GPT, Claude, Gemini, Grok, DeepSeek, Qwen | Cursor, Perplexity, Higgsfield, most "AI" apps you see |
Some companies do both. OpenAI makes GPT models and the ChatGPT app. Anthropic makes Claude models and the Claude app. Runway (video), ElevenLabs (voice), and Suno (music) also build their own models and their own apps.
Most companies only build the harness. They rent a model from one of the labs and focus on making it useful for one job. When people say "AI agent" or "agentic product," they usually mean a harness that gives a model a task, lets it use tools, and checks its work step by step.
Who makes the models (October 2026 snapshot)
As of October 2026, Anthropic and OpenAI hold the top of the rankings, and Google sits just behind them. Rankings change almost monthly, so treat this as a photo, not a map.
| Group | Makers | Where they stand | What they compete on |
|---|---|---|---|
| Frontier | Anthropic (Claude Opus, Sonnet, Fable), OpenAI (GPT-6 Astra, GPT-6.1 Sol) | Every spot in BenchLM's overall top 10 belongs to these two plus one Google model | Best quality, highest prices |
| Close behind | Google (Gemini 4 Argon) | One model in the top 10 | Long documents, images and video input, mid-range prices |
| One step back | DeepSeek, Qwen (Alibaba), Kimi (Moonshot), GLM (Zhipu) | Best Chinese model on BenchLM ranks about 17th | Very low prices; many are "open-weight," meaning anyone can download and run them |
| Same band, pricier | xAI (Grok), Meta | Below the frontier, near the Chinese models | Grok sits mid-price; Meta has moved to closed models |
Video, voice, and music have their own leaders. In video, Google's Veo 3.1 is the common all-round pick, while two Chinese models, Kling 3.0 and Seedance 2.0, lead several blind-vote rankings at lower prices. Midjourney and Runway are favorites for artistic control. OpenAI shut down its Sora video app in 2026. For voice, ElevenLabs is the best-known name; for music, Suno.
A note on how the cheaper models catch up: some labs are widely reported to train smaller models on answers from frontier models. This is called distillation. It is one reason the gap stays at about one step.
Sources: BenchLM frontier rankings (https://benchlm.ai/frontier-ai-models (opens in a new tab)), Techsy video model rankings (https://techsy.io/en/blog/best-ai-video-models (opens in a new tab))
What it costs
For everyday chatting, you pay a flat monthly subscription or use a free tier, and cost barely matters. Cost matters when you build something that runs a model thousands of times. Then the gap between models is about 40 times.
Models charge by the token. A token is a chunk of text, roughly three-quarters of a word. Prices are quoted per million tokens, with separate prices for what you send in (input) and what the model writes back (output).
| Model | Maker | Input, $ per million tokens | Output, $ per million tokens | Example month: 50M in + 10M out |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10.00 | $50.00 | $1,000 |
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | $1,000 |
| Claude Opus 5.5 | Anthropic | $4.00 | $20.00 | $400 |
| Gemini 4 Argon | $2.00* | $10.00* | $200* | |
| GPT-6.1 Sol | OpenAI | $2.00 | $10.00 | $200 |
| Grok 4.7 | xAI | $2.00 | $6.00 | $160 |
| DeepSeek V4 Pro | DeepSeek | $1.32 | $3.96 | $106 |
| DeepSeek V4.1 Flash | DeepSeek | $0.30 | $1.20 | $27 |
* Google's Gemini 4 Argon price doubles on December 1, 2026, to $4.00 in and $20.00 out ($400 for the example month).
The example month is a small business tool processing about 37 million words of documents. On the top model it costs about $1,000. On the cheapest Chinese model it costs about $15 to $30. That is why heavy, repetitive work usually runs on cheap models, and the expensive ones are saved for hard thinking.
Prices are standard list prices, rounded, as of October 23, 2026, for prompts under each provider's long-context threshold. DeepSeek charges half these rates off-peak. Discounts for bulk ("batch") work and repeated text ("caching") can cut them by half or more. Sources: official pricing pages from Anthropic, OpenAI (opens in a new tab), Google (opens in a new tab), xAI (opens in a new tab) and DeepSeek (opens in a new tab).
Harnesses you may already use
Most AI products are harnesses, and they fall into a few types. Many let you pick which model runs underneath.
| Type | What it does | Examples |
|---|---|---|
| General chat | A chat window for any question | ChatGPT, Claude, Gemini, Grok, plus many look-alike apps |
| Coding | Writes and edits software inside your project | Cursor, Claude Code, GitHub Copilot |
| Search | Searches the web and writes an answer with links | Perplexity |
| Video and images | Sends your idea to one or more video or image models | Higgsfield (runs Veo, Kling, and Seedance) |
| Agents | Takes a goal and works through the steps on its own | Assistants that read email, book things, or run sales tasks |
A quick test: if a product lets you switch between models from different companies, it is a harness. Cursor, for example, can run models from Anthropic, OpenAI, and Google.
The model is becoming the swappable part
Models keep getting better and cheaper, and a new one appears every few weeks. So the model is starting to act like a power outlet: important, but replaceable. What makes a tool worth using is what it builds on top.
A good harness owns things a bare model cannot do alone:
- A method. A tested way of working through a problem, step by step, instead of one big question.
- Your context. What it has learned about you, your work, and your past projects.
- Its data. Records, examples, and results that build up as people use it.
- Its connections. Access to your email, files, calendar, or other apps, with your permission.
The weak spot: a harness that only sends your words to a model, with nothing added, is easy to copy and easy to replace. When you pick a tool, ask what it does that the plain chat app does not.
What this means for you: learn the habits, not the brand. If you know how to ask well, check the answer, and give good context, you can move to whatever tool is best next year.
AI executes, humans create
This is the idea behind the whole series. AI is a tool. It doesn't replace you; it extends what you can do, the same way a calculator didn't replace people who are good with numbers.
AI has already read a huge share of everything humanity has written, and that share grows every year. That means anything AI can produce by rearranging what already exists isn't new. If AI can do it, it's not the best use of your energy.
What AI can't produce is the genuinely new: ideas from your own life, things you noticed firsthand, solutions to problems no one has written about yet. That is your job. Let AI handle the execution, so you can spend your time creating what doesn't exist yet.
For anyone whose sense of self is built around getting tasks done, this can be a wake-up moment. It's a chance to decide whether it's time to move from doing the work to deciding what work is worth doing.
Words to know
| Word | Plain meaning |
|---|---|
| Model (LLM) | The AI brain that reads and writes text. LLM means "large language model." |
| Harness | The app or tool built around a model. |
| Agent | A harness that takes a goal and works through steps on its own. |
| Token | A chunk of text, about three-quarters of a word. How usage is measured and billed. |
| Frontier model | One of the best models available right now. |
| Open-weight | A model anyone can download and run on their own computer. |
| Distillation | Training a smaller model on a bigger model's answers. |
| Prompt | What you type to the AI. |
Found this helpful? Raya.ai is built on the ideas in this series. Learn more at raya.ai (opens in a new tab), or download the app on iOS (opens in a new tab) or Android (opens in a new tab).