The gap is closing

On LMArena (Chatbot Arena) the best closed model scores 1501 and the best open-weight model scores 1470. A gap of 31 Elo, down from more than 100. Capability is commoditising. Running it still costs money.

Each building is a model, climbing as it wins its matches.

In 2023, the best of them lived behind a wall.

Then the open models started to climb.

By 2025 they had drawn level.

Today the wall is a step.

But every one of these has to be powered.

Every AI model

competes in an arena.

Two models answer the same prompt.

A person picks the better reply.

Win, and you climb.

Lose, and you fall.

The score is called Elo.

The same idea behind chess rankings.

Three years of that contest,

laid out as a city.

Scroll

The smartest model stopped being the moat.

When every model is good enough, the question is no longer which one.

It is what it costs to keep them running.

What you just watched

Every building is an AI model. Slate is closed, sage is open weights. Height is its Elo.

Elo works like chess ranking. Two models answer the same prompt, a person picks the better reply, and the winner takes points from the loser. Millions of these votes on LMArena (Chatbot Arena) set every score. Over three years the open skyline climbed to within 31 Elo of the closed one.

Capability is becoming a commodity. The next constraint is energy. Someone has to power every one of these, and that bill only grows.

Knowing what your AI actually costs to run is the problem I work on at flowstate. Data: LMArena (Chatbot Arena) overall leaderboard, snapshot 2026-06-16.

The frontier, in numbers

The tallest closed tower reaches 1511 Elo, the tallest open-weight tower 1478. A gap of 33, down from more than 100 in early 2024.

Metric: LMArena (Chatbot Arena) Arena Elo, overall category. Each building is one model, its height the peak Elo it reached.
Closed: Proprietary license.
Open: Open-weight license.
Window: 8 May 2023 to 16 June 2026.

Source: LMArena (Chatbot Arena) leaderboard, built from the lmarena-ai dataset on Hugging Face.

Back to willhackett.com

The closing AI frontier

In 2023 the best closed model led the best open-weight model by a wide Elo margin. By 2026-06-16 the best closed model scored about 1511 and the best open-weight model about 1478 on LMArena (Chatbot Arena), a gap of about 33 Elo, down from more than 100. The capability moat has nearly collapsed. The remaining differentiator is the cost to run these models.