← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

The AI Model
Cheat Sheet

The 8 AI models worth knowing right now, what each one is best at, what to watch out for, and what it costs. Plus the switch-between-models workflow that beats picking a favorite.

Everyone asks me which AI model is the best. I think that’s the wrong question. The better one is which model is best for this task.

The people getting the most out of AI stopped picking a favorite. They switch models based on the job in front of them. Here’s the lineup I’d know today, and the workflow that ties it together. Save this one.

The 8 Models

Costs are API prices per million tokens, checked November 1, 2026 on each company’s own docs (Anthropic, OpenAI, Google, xAI). In the chat apps you pay a monthly plan instead, but the price tells you which models are heavy.

01

Claude Fable 5.1

Anthropic · $10 / $50 per million tokens in / out

Reach for it when: you have the hardest reasoning problem or a long agent job, and another model already stalled on it.

Watch out: it’s the slowest and priciest Claude. On the Pro plan it runs on usage credits.

02

Claude Opus 5.5

Anthropic · $4 / $20 per million tokens in / out

Reach for it when: you want one strong model for coding, research and long knowledge work. Anthropic itself says to start here for most work. It’s also my pick for writing people will actually read.

Watch out: overkill for quick back and forth.

03

Claude Sonnet 5.5

Anthropic · $2 / $10 per million tokens in / out

Reach for it when: you want speed and smarts in one. My daily driver for everyday drafting, coding and documents.

Watch out: step up to Opus or Fable when it starts missing the hard parts.

04

GPT-6.1 Sol

OpenAI · $2 / $10 per million tokens in / out

Reach for it when: you’re doing complex coding, computer use or professional work and want near-top results for less. OpenAI’s own top tier, GPT-6 Astra ($10 / $50), sits above it.

Watch out: in ChatGPT, your plan decides which tier you get when you pick GPT-6.

05

GPT-6 Luna

OpenAI · $0.10 / $0.50 per million tokens in / out

Reach for it when: you have a pile of small repetitive jobs to run cheaply, like tagging thousands of rows.

Watch out: it’s built for focused tasks, so check its work on anything with stakes.

06

Gemini 4 Argon

Google · $2 / $10 per million tokens in / out

Reach for it when: you want Google’s flagship reasoning model, or your work already lives in Google Workspace.

Watch out: Google’s list price rises to $4 / $20 on December 1, 2026.

07

Gemini 3.8 Flash

Google · $0.75 / $3.75 per million tokens in / out

Reach for it when: you want speed at a low price. Google built it for long software jobs, agents and complex workflows.

Watch out: that price is introductory and doubles on January 1, 2027.

08

Grok 4.7

xAI · $2 / $6 per million tokens in / out

Reach for it when: you need to know what happened in the last few hours. Grok can search X posts and threads live.

Watch out: prompts of 200,000 tokens or more are billed at double the rate.

If You Only Remember Four Things

Hardest problem: Claude Fable 5.1. Everyday work: Claude Sonnet 5.5 or GPT-6.1 Sol. Cheap and high volume: GPT-6 Luna or Gemini 3.8 Flash. What happened in the last hour: Grok 4.7.

There’s also a free lane. Open-weight models like DeepSeek, Qwen, Kimi, GLM and Google’s Gemma can be downloaded and run yourself. Open weight means you get the finished model but not the training data or recipe.

The Three-Model Workflow

The workflow I keep seeing has three steps, and you don’t use one model for all of them.

01

Think

Start with your smartest model to plan the strategy and break the problem into pieces. You want deep reasoning here, not speed. In my flow that’s Claude Fable 5.1 or Opus 5.5.

The Planner
You are my planning partner. I'm going to hand the build to a different, faster model afterward, so your job is the thinking, not the doing. The goal: [what you want to end up with, e.g. "a launch email sequence for my new course"] What I already have: [paste or attach notes, data, past examples] Constraints: [deadline, budget, tone, tools I use, anything off limits] Do this: 1. Ask me up to 3 questions if something important is missing. Wait for my answers. 2. Break the job into clear steps, in order. For each step, say what "done" looks like. 3. Flag the 2 riskiest assumptions in the plan and how to test them. 4. Write a handoff brief I can paste into another model to do the build: the goal, the steps, the inputs, the exact output format, and what to avoid. Rules: think it all the way through before you answer. Don't start building. If something can't work, tell me plainly.

02

Build

The hard thinking is done, so switch to a faster, cheaper model and paste in the handoff brief from step one. Sonnet 5.5 or GPT-6.1 Sol both do the heavy lifting well.

03

Review

The step most people skip. Hand the finished work to a different model to find mistakes. A model that built something tends to read it the same way it wrote it, so it misses its own gaps. A fresh model reads it cold, the way a friend catches the typo you read past ten times.

The Skeptical Reviewer
You are a skeptical reviewer. A different AI model made the work below. You didn't make it, so you have no reason to defend it. Your only job is to find what's wrong before I use it. What it was supposed to do: [paste the original goal or brief] The work: [paste or attach the finished draft, code, plan or analysis] Who it's for: [audience and what they'll do with it] Check, in this order: 1. Does it actually do what the brief asked? List anything missing. 2. Facts, numbers and claims: flag anything wrong, unsupported or likely made up. Say how I can check each one. 3. Logic: gaps, contradictions, steps that don't follow, edge cases it ignores. 4. For code: bugs, missing error handling, anything that would break on real data. 5. Clarity: the 3 spots a reader is most likely to get confused. Output: a numbered list, most serious problem first. For each one, say where it is, why it matters, and the exact fix. End with a one-line verdict: ready, ready after fixes, or start over. Rules: don't rewrite the whole thing. Don't praise it. If something is fine, skip it.

The One Rule

The reviewer has to be a different model than the one that did the work. Same model, same blind spots. Switch the model and you switch the point of view.

One Honest Note

Don’t get attached to the exact names. Every model on the July version of this list has already been replaced, and it’ll happen again. What stays true is the shape. Use your smartest model to think and a faster one to build. Get a different one to review, and send the boring stuff to the cheapest.

I think of AI models like streaming services. If you only ever open one, you miss the good stuff on the others. The advantage isn’t owning the one perfect model. It’s knowing which one to open for the job in front of you.