The Three Tiers of the Market
The AI industry looks chaotic from the outside because there's a new model announcement roughly every day. But if you zoom out, there are really only three tiers of players that matter to a working professional, and every tool you encounter fits neatly into one of them.
Tier 1: the frontier labs. These are the handful of companies actually training the biggest, most capable models in the world from scratch. They're the only ones running the multi-billion-dollar training runs. In 2026 the short list is Anthropic, OpenAI, and Google DeepMind. Everyone else in this tier is either spending a lot of money to catch up or has quietly dropped out of the race. Your chatbot is almost certainly powered by one of these three, whether you realize it or not.
Tier 2: open-weight models. Models whose weights — the internals — have been released publicly so anyone can download them and run them on their own computers or servers. Meta's Llama family is the most famous example, but there's a long tail of high-quality open releases from Mistral, DeepSeek, Qwen, and others. These models trail the frontier by a few months to a year in raw capability, but they're free, private, and you can run them without sending data to anyone.
Tier 3: the wrappers. Every other AI tool you see in the App Store, every “AI-powered” product inside an existing SaaS, every chatbot built by a startup. Almost none of them are training their own models. They're calling one of the Tier 1 or Tier 2 models via an API, wrapping it in a product, and selling the wrapper. That's not a criticism — a lot of the real value comes from good wrappers — but it means when you evaluate a tool, the first question is "whose model is under the hood," and the answer is almost always a name from Tier 1.
Hold those three tiers in your head and the market stops looking chaotic. Every product is either a frontier lab, an open model, or a wrapper. That's it.
Anthropic — Claude
Anthropic is the company behind Claude. Founded in 2021 by Dario and Daniela Amodei and a group of researchers who previously worked at OpenAI, it positions itself as the “safety-first” frontier lab. That positioning isn't just marketing — the company's research output on interpretability, alignment, and responsible deployment genuinely leads the industry, and it shapes how Claude behaves.
What Claude is known for: being unusually good at careful writing, long-document reasoning, and coding. It's the model most programmers I know reach for when they want an AI that reads a whole codebase carefully before it acts. It's also the model most writers I know reach for when they want something that can match a voice and push back on weak arguments instead of flattering them. That willingness to push back — to say “I don't think you should do it this way” when you're about to do the wrong thing — is the single most distinctive thing about Claude and the reason I use it for anything I care about.
Where you'll encounter it: the Claude.ai web app, the Claude mobile app, Claude Code (the coding agent I use to maintain this site), and as the model behind a lot of enterprise AI products. Amazon, through its partnership with Anthropic, makes Claude available inside AWS for companies that want to self-host their AI workloads.
When I reach for it: anything that involves long documents, writing that needs to sound like me, code that touches a project I care about, or decisions where I want a second opinion from something that will disagree with me if I'm wrong.
OpenAI — ChatGPT and GPT
OpenAI is the company that started the current era. ChatGPT's launch in late 2022 was the moment this all went mainstream, and the company has been the most commercially dominant frontier lab ever since. OpenAI trains the GPT family of models and runs ChatGPT, which remains the most-used AI product in the world by a wide margin.
What GPT is known for: breadth. ChatGPT is excellent at a huge range of tasks and is usually among the best at the one you happen to need. The company ships fast and incorporates new capabilities — image generation, voice, browsing, code execution, custom agents — into the chat interface faster than anyone else. If you want one AI tool and you don't want to overthink it, ChatGPT is still the default recommendation for a huge fraction of users.
Where you'll encounter it: ChatGPT, of course. Also inside Microsoft 365 Copilot (Microsoft is OpenAI's largest investor and distributes GPT models throughout its Office and Azure products). Also under the hood of countless third-party tools — if you see "AI-powered" in a product and can't figure out whose model it is, there's a decent chance it's the OpenAI API.
When I reach for it: when I need image generation alongside text, when I want voice mode, when I'm experimenting with a new feature OpenAI shipped last week, or when someone I'm helping already uses ChatGPT and I'm meeting them where they are.
Google — Gemini
Google's family of models is called Gemini, and the company has been closing the gap with the other two frontier labs steadily for the last two years. Gemini is trained and operated by Google DeepMind, which merged DeepMind and the old Google Brain into a single research org back in 2023.
What Gemini is known for: massive context windows, strong multimodal capabilities (text, image, audio, video all in the same model), deep integration with the rest of the Google ecosystem (Search, Workspace, Android), and raw speed on big tasks. Gemini can comfortably handle prompts that include a whole book, a whole video, or a whole codebase in one pass, which is a real differentiator for research and analysis work.
Where you'll encounter it: gemini.google.com directly. Also inside Google Workspace (Docs, Sheets, Gmail) for paying subscribers. Also as the model powering NotebookLM, the Google product that turns a pile of source material into a podcast-style audio summary — probably the single most widely-used AI product that casual users don't realize is Google. On Android phones Gemini is increasingly baked into the operating system itself.
When I reach for it: when I have an enormous document or a multi-hour video to analyze, when I'm already working inside Google Docs or Sheets, when I want image generation that feels different from OpenAI's, or when I need NotebookLM to build a training video from a set of sources.
Meta and Open Source — Llama and Friends
Meta (the company that owns Facebook, Instagram, and WhatsApp) doesn't sell access to its models the way Anthropic or OpenAI do. Instead, it trains large models called Llama and releases the weights publicly so anyone can download them. This is the cornerstone of the open-source AI ecosystem.
Why open models matter: three reasons. First, price — once you have the weights, running the model is just compute cost, no per-token API fees. Second, privacy — the data you feed an open model you run yourself never leaves your machine or your company's servers, which matters a lot for regulated industries. Third, customization — you can fine-tune an open model on your own data to make it much better at your specific task, which you can't easily do with a closed API model.
Other players in the open-weight tier: Mistral (a French lab that makes small, fast, capable models), DeepSeek (a Chinese lab whose 2025 releases genuinely rattled the frontier), Qwen (from Alibaba), and a long tail of research releases from universities and smaller labs. The open-weight ecosystem is usually a few months behind the very best closed models, but the gap has been narrowing.
When it's worth the detour: if you run a business with sensitive data and a competent technical team, if you want to build a product without a per-request bill attached, or if you need to fine-tune on proprietary data. For everyday conversational use, the closed Tier 1 models are still easier and more capable out of the box.
Microsoft — The Distribution Layer
Microsoft doesn't fit cleanly into "frontier lab" or "wrapper" because it's doing something more interesting: it's the distribution layer for almost everything else. Microsoft has a multi-billion-dollar strategic partnership with OpenAI, runs the Azure OpenAI Service (which is how most Fortune 500 companies consume GPT models), owns GitHub, and ships Copilot throughout Microsoft 365.
What Copilot actually is: not one product but a family of them. Copilot in Word, Copilot in Excel, Copilot in Outlook, Copilot in Teams, Copilot in PowerPoint, Copilot in Windows, GitHub Copilot for code. Each one is a different AI feature using (usually) OpenAI models, deeply integrated with the product it lives inside. The deep integration is the whole point — Copilot in Outlook can read your actual emails, Copilot in Excel can work with your actual spreadsheet, Copilot in Word can edit your actual document. That contextual access is something no general chatbot can match.
Why it matters for working professionals: if your company pays for Microsoft 365, you probably have some form of Copilot already licensed, whether you've used it or not. And for a lot of office tasks — drafting emails, summarizing meetings, extracting data from spreadsheets — Copilot is better than a separate chat tool precisely because it lives inside the application where your work already is.
For the full practical breakdown of Copilot at work, see the Copilot guide in the Current Tools pillar.
How to Actually Pick a Tool
The question most working professionals ask is "which AI should I use?" That's the wrong question. The better version is "which AI should I use for this task, and am I willing to keep two or three of them around for different kinds of work?" The honest answer for most people is yes.
Here's the practical rule of thumb I use:
- If your task lives inside Microsoft 365 (drafting email in Outlook, summarizing a Teams meeting, analyzing an Excel file), start with Copilot. The contextual access to your actual files is worth more than any model-quality difference.
- If your task is long-form writing, careful reasoning, or coding you care about, reach for Claude. It's the best at not flattering you into bad decisions.
- If your task is experimental, multimodal, or you just want the fastest path to "a usable answer", use ChatGPT. It's the generalist with the biggest toolbox.
- If your task involves a huge document, a long video, or you're already in Google Docs or Sheets, try Gemini. The massive context window is a genuine advantage.
- If your data cannot leave your company's servers, look at open-weight models — Llama, Mistral, DeepSeek — and plan on a technical team to run them well.
Most professionals I know keep two or three of these tools open and switch based on the task. That sounds like overhead, but after a few weeks the switching becomes automatic and the gains are real. Treat them like tools in a toolbox. You wouldn't use a hammer for every job.
What Actually Changes Week to Week (Spoiler: Less Than You Think)
If you read AI news, the industry sounds like it's in constant upheaval — new model every week, new leader every month, new capability every day. Most of that is noise that doesn't affect how you use these tools.
Here's what actually matters to a working professional, and how often:
- New frontier model releases from the Tier 1 labs: a few times a year per lab. Each one is usually smarter than the last in predictable ways. Worth noting; not worth chasing in real time.
- New features in the chat products (voice, image, browsing, agents, memory): a few times a month across all three. Worth trying when they land. Most don't change your daily workflow.
- New open-weight model releases: weekly. Relevant only if you're running open models yourself. Otherwise, ignore.
- The underlying physics: hasn't meaningfully changed since 2022. Models predict the next token. They hallucinate. They forget. They have context limits. The four words in Module 2 still explain 90% of the weird behavior you'll observe.
The practical rule: check the state of the field every two or three months. Otherwise, pick a setup that works and stop chasing the weekly news cycle. The frontier labs aren't going anywhere; your chosen tool isn't going to be abandoned tomorrow; and the one-week difference between the best model and the second-best model almost never matters for the real work you're doing.
The landscape is stable enough that learning it once — who's who, what each lab is known for, which tool to reach for when — gives you a map that's still useful three months later. That's the whole point of this module. You now have the map. Stop refreshing the news feed. Go do the work.