How to read it

The index

Intelligence numbers are the Artificial Analysis Intelligence Index, read from their LLM leaderboard on 1 Oct 2026. The leaders sit in the mid-50s. A 54 is not 54%. The bars on the board run to 60 so a frontier score looks like a frontier score.

Effort

The same model at low and at max can be ten points apart, and several dollars apart per task. The control at the top asks every model for the same setting. If that setting was not published, the card shows the closest one and says so. Recommendations prefer a real score at the setting you picked.

Two different prices

“Per measured task” is what Artificial Analysis spent to run that model through one of its own index tasks, reasoning tokens included. It is the right number for “how expensive is this model when it thinks.” It is the wrong number for a chatbot reply. The calculator uses the lab’s list price on a workload you can edit.

Blended price is three input tokens plus one output token, the mix Artificial Analysis uses, divided by four. It is a sticker comparison. Verbose models look cheap on the sticker and expensive per task.

Where the prices came from

OpenAI, Anthropic, Google Gemini 3.8 Flash, and SpaceXAI prices were read from those labs’ own pricing pages. Gemini 4 Argon’s price is the introductory rate in Google’s launch post. It was not on the public pricing page, because the model was not generally available. Open-weight token prices are vendor list prices compiled on 23 Sep 2026, except DeepSeek’s peak and off-peak rates, which come from its API pricing write-ups. Where a yuan conversion disagrees by a few cents, the model card says so.

What was left out

Arena Elo, vendor-run SWE tables, and one-off launch slides are not columns. They answer different questions and they contradict each other often enough to make a merged score worse. A few Artificial Analysis evals that were published in the same sentence, such as Terminal-Bench 4.0, are on the model page and labeled.

Mistral, Amazon, Cohere, NVIDIA, IBM, and MiniMax are named on the labs page. They did not have a model in the frontier cut used here. An empty cell means the number was not in the sources for this snapshot. It is not a zero.

The frontier line

A model is on the frontier when no other model in the current list is both at least as strong and strictly cheaper per measured task, or both cheaper and strictly stronger. Change the effort or the filters and the line changes. That is the point.