A non-developer's map of the AI model landscape: the major makers, what each model is for, and how to pick one. Companion to the AI & Claude Glossary.
Pick in 30 Seconds
Every maker offers the same three-tier shape. Learn the shape once and it transfers everywhere:
| Top tier "the smart one" | Most capable, slowest, most resource-heavy. For hard reasoning, strategy, nuanced writing, high-stakes work. Claude Opus / Fable, GPT-5, Gemini Pro. |
| Mid tier "the workhorse" | The balanced everyday default — fast enough, smart enough. 90% of work lives here. Claude Sonnet, GPT-5 mini, Gemini Flash. |
| Light tier "the quick one" | Fastest and lightest. For simple, high-volume tasks — tagging, sorting, short replies. Claude Haiku, GPT-5 nano, Gemini Flash-Lite. |
CMO rule of thumb: start on the mid tier. Move up when quality visibly matters (a board memo, a positioning doc); move down when you're doing the same simple thing thousands of times (classifying inbound leads). You rarely need the top tier as your default.
Claude — Anthropic (what you're using)
Anthropic's lineup and what each is for. "Context" = how much it can read at once; "1M tokens" ≈ 750,000 words (a long book).
| Model | What it's for | Context |
|---|---|---|
| Fable 5 | The most powerful tier — the hardest, highest-stakes reasoning | 1M |
| Opus 4.8 | Top everyday "smart one" — strategy, deep analysis, long autonomous tasks | 1M |
| Sonnet 4.6 | The balanced workhorse — daily drafting, summarizing, most work | 1M |
| Haiku 4.5 | Fastest & lightest — quick classifications, simple high-volume tasks | 200K |
(Opus 4.7 and 4.6 are still available as slightly older versions of the "smart one" — you'd only pick them deliberately.)
Why matching matters: a quick classification doesn't need the most powerful model, and a board memo shouldn't run on the lightest one. The discipline is "use the lightest model that's good enough" — it's faster and keeps the heavy models free for the work that truly needs them.
The Other Big Makers
The companies a CMO will actually hear named, and what each is known for. Version numbers shift constantly — these describe the lineup and reputation, not a spec sheet.
| Maker | Model family | Known for |
|---|---|---|
| Amazon | Nova — plus Bedrock, which hosts Claude, Llama, etc. | AWS-native; Bedrock is how many enterprises access many models through one contract. |
| Anthropic | Claude (Opus / Sonnet / Haiku, Fable) | Strong reasoning & writing, safety focus, long context, enterprise trust. (What HueLife uses.) |
| DeepSeek | DeepSeek (open-weight) | Strong reasoning for very little compute; Chinese lab (raises some data-governance questions for enterprises). |
| Gemini (Pro / Flash / Flash-Lite) — in Workspace & Search | Massive context, strong multimodal (image/video/audio), deep Google-apps integration. | |
| Meta | Llama | Open-weight (free to download & self-host). Popular when you want control or to avoid per-use fees. |
| Microsoft | Copilot (runs OpenAI models) + small "Phi" models | Baked into Office/365, Windows, GitHub. Likely your org's most familiar AI surface. |
| Mistral | Mistral / Mixtral | European (EU-data-friendly), efficient, several open-weight options. |
| OpenAI | GPT-5 family; "o-series" reasoning models — powers ChatGPT | The household name, biggest ecosystem, image generation, broadest plugin/app support. |
| Perplexity | An answer engine (uses others' models under the hood) | AI web search with citations — a research tool, not a model maker. |
| xAI | Grok — tied to X / Twitter | Real-time access to X data; positioned as less filtered. |
The one distinction worth knowing: closed/hosted models (Claude, GPT, Gemini) you rent through an account — easy, no infrastructure. Open-weight models (Llama, DeepSeek, Mistral) you can download and run yourself — more control and privacy, but you (or a vendor) must host them. For a marketing team, hosted is almost always the right call.
Decoding Model Names
Model names look like alphabet soup. Here's how to read them:
| A bigger number (4.6 → 4.8, GPT-4 → GPT-5) | Newer generation = generally smarter. Usually the only number that matters. |
| "mini" / "nano" / "Flash" / "Lite" / "Haiku" | A smaller, faster, cheaper version of that generation — same family, less horsepower. |
| "Pro" / "Opus" / "Ultra" | The bigger, more capable version — more horsepower, higher cost. |
| "reasoning" / "thinking" model | A mode (or model) that works through problems step-by-step before answering. Slower, better at hard logic. (Claude does this with its adaptive thinking + effort settings.) |
| Context window (128K, 200K, 1M) | How much it can read at once. Bigger = can digest longer documents without forgetting the start. |
| Multimodal | Handles more than text — images, PDFs, sometimes audio/video. Most current top models are multimodal. |
| Open-weight vs. closed | Open = downloadable/self-hostable (Llama, Mistral, DeepSeek). Closed = rented via an account (Claude, GPT, Gemini). |
Which Model for Which Job (CMO edition)
| The task | Reach for… |
|---|---|
| Board memo, positioning doc, nuanced strategy | Top tier — Claude Opus 4.8 or Fable 5 |
| Everyday drafting, email, summaries, campaign copy | Mid tier — Claude Sonnet 4.6 (your default) |
| Classifying 5,000 inbound leads, tagging, sorting | Light tier — Claude Haiku 4.5 |
| "What's the latest on X / who is this company?" | A web-grounded tool — Perplexity, or Claude/ChatGPT/Gemini with web search on |
| Reading a 100-page report or contract in one go | Any large-context model — Claude (1M), Gemini (huge context) |
| Inside Office/Outlook/Teams | Microsoft Copilot (it's already there) |
| Inside Google Docs/Gmail | Google Gemini (built into Workspace) |
The one-paragraph version: every AI company sells the same three tiers — a smart one, a workhorse, and a quick one. Pick the cheapest tier that's good enough, default to the workhorse, and only reach for the top tier when quality is visible and stakes are high. Claude (Anthropic) is your house model; ChatGPT (OpenAI) and Gemini (Google) are the other two giants; Copilot is Claude/GPT power baked into Microsoft 365. Everything else is a variation on those ideas — and the version numbers will have changed by the time you finish reading this.