What's the largest number of player avatars ever in one area of a virtual world or MMO?

“One area” means they can all see and potentially interact with each other. Sharding doesn't count. Only more or less human avatars moving around at ground level count. NPCs don't count. FPS down to unplayable levels doesn't count.

Improbable and Dual Universe both did “stress tests” with numbers around 10,000, but they were just spaceships zipping around. Improbable claims their architecture can do big crowds, but they've never shown it.

The record for movies is around 200,000, in "Lord of the Rings" (2001). Impressive for 20 years ago.

That's the wrong question to ask. It all depends on factors such as how big the characters are on-screen, how detailed they are, how much bandwidth they use, and so on.

For example, a typical full HD screen has 2073600 pixels. Each of these pixels could represent another player. They couldn't move (there's no room to move; the screen is packed full), but they could, for example, change the color of their pixel. Not all at once, probably, but each of them every few seconds or so. That's the upper limit if you want to sacrifice everything else to put the maximum number of distinct individual characters on the screen at once.

Or at the other extreme, maybe you want super-fancy 3D avatars with no compromise on visual quality? You can only have one of those on the screen at a time, since anything else on the screen by definition requires diverting some GPU resources away from that one 3D avatar to whatever else is being rendered on the screen.

The Lord of the Rings was absolutely “unplayable FPS” – it was rendered offline, with render times of “hours per frame.”

The question doesn't make a lot of sense, though, because it sounds a bit like “how long is a piece of string?"

The cost of animating characters matters. If a character is a single 2D sprite, you can draw very many of them. If a character is a 50,000 polygon avatar with 150 animated bones and full PBR shading including raytraced reflections, you can render many fewer of them.

If the avatars have no collision, you can put many of them in one place. If they have simple ("sphere" or “capsule”) collision, you can put fewer of them in one place, but still a fair number, especially if you accept penalty-based methods and sudden “ejections” when too many characters squeeze together. If you fully physically simulate each leg/arm/body/head of each avatar, and don't accept non-physical behavior, the limit is going to be much lower.

And, finally, the question also depends on the hardware. A workstation Threadripper with a RTX 3090 graphics card running highly tuned C++ and CUDA and D3D12 code will do differently than a dual-core Intel Pentium laptop with Intel Integrated graphics on a javascript webpage game.

So, the answer is: “Between 10 and 1,000,000.”