Skip to the content.
Reference This entry is primarily explanatory reference: what the term means, why it exists, and how it is used.

Closed source

The other large strategic camp in the contemporary AI model ecosystem.


In the Elon Musk / Patrick Collison / Dwarkesh Patel interview, Musk pauses on the names of the AI labs and enjoys the irony: “Midjourney is not mid. Stability AI is unstable. OpenAI is closed. Anthropic? Misanthropic.”1

The joke lands unevenly. OpenAI really is closed in the sense that matters for this entry: GPT-5 is not a file you can inspect, download, or run. Anthropic is not therefore misanthropic; that is a clever line, not a fair diagnosis. And the contrast matters because there really are models you can download and run locally — Gemma 4 and Qwen, for example, sit on the open-weights side of the ledger. The names are funny. The architecture is serious.

In one sentence

A closed-source AI model is one whose weights are kept proprietary; users access the model only through an API hosted by the vendor, who charges per token, controls the training and update cycle, and sees every prompt that flows through.

Why this term exists

By default, building a frontier AI model is enormously expensive — hundreds of millions of dollars in compute alone, plus the equivalent in research talent and data acquisition. A lab that has spent that much wants to recoup the investment, and the most direct way is to keep the model proprietary, expose it as an API, and charge per use.

The closed-source camp in mid-2026 includes the most capable models commercially available: OpenAI’s GPT-5, Anthropic’s Claude (Opus, Sonnet, Haiku), Google’s Gemini Ultra and Pro, and most of the next tier of well-funded contenders. The open-weights camp includes models such as Gemma 4 and Qwen, which can be downloaded and run under local or self-hosted arrangements. The frontier-capability ceiling, in 2026, is still held by closed models. The open-weights ecosystem is closing the gap, but is not yet at parity.

What it actually does — concretely

A closed-source model is delivered as a service, not a file. You:

  1. Sign up for an account with the vendor.
  2. Get an API key.
  3. Send HTTP requests with your prompts.
  4. Receive responses.
  5. Pay for what you used at the end of the month.

You do not have access to the model weights, the training data, the exact architecture (sometimes a rough description is published in a paper, sometimes not), or any guarantee that the model you call today will behave identically to the model you called last month.

Working example

The OpenClaw setup on this MacBook calls Anthropic’s Claude Opus 4.7 as its primary model. The interaction looks like:

Your laptop  →  HTTPS POST  →  Anthropic's data center  →  GPU cluster
                                                            running Opus 4.7
                                                                  ↓
                                                              response
                                                                  ↓
Your laptop  ←  HTTPS reply  ←  Anthropic's data center  ←  GPU cluster

The model itself is on Anthropic’s hardware. Your prompt and response transit Anthropic’s network. The billing meter ticks for every token in either direction. The capability is real and excellent; the dependence is also real and total.

When Anthropic raises prices, deprecates a model, changes terms of service, or has a regional outage, every workflow you have built on Opus is affected. This is not a criticism of Anthropic specifically — it is the structural reality of closed-source.

Why this matters in a teaching context

For a BBA or MBA classroom, the open-vs-closed AI conversation is a clean case of commoditize-your-complement strategy taught through current events.

Worth working through with students:

The strategic positions of these players, mapped against the open-vs-closed axis, are some of the cleanest competitive-strategy material currently available for an Isenberg classroom — especially when paired with the equipment-auctions research thread, which has the same shape: who profits when the underlying asset commoditizes?

Trade-offs


Related entries: Open source, Token burn, Token angst, Ollama.

  1. Elon Musk in conversation with Patrick Collison and Dwarkesh Patel, May 2026, cleaned transcript on file in the Dictionary workspace as Interview with Elon Musk.md. The exact line in the cleaned transcript: “Midjourney is not mid. Stability AI is unstable. OpenAI is closed. Anthropic? Misanthropic.” This entry accepts the OpenAI part of the joke as structurally accurate and rejects the Anthropic part as too glib. 

Return to Dictionary All Entries (A–Z) For Students Other Writing