For two years the intelligence index was a two-horse race. In eight days this July, the frontier became a street.
AiSignal staff · 4 September 2026 · 7 min read
Artificial Analysis keeps the only leaderboard that product people actually argue about. In early June, two labs had a model above 50 on the Intelligence Index. By mid-July there were six: Anthropic, OpenAI, Moonshot, xAI, Z.AI, Meta. Four of those launches landed in eight days. The top three scores came from three continents and sat three points apart.
September's v4.2 print is already a different poster. Claude Fable 5.1 leads. GPT-6 Astra took a four-point jump on Sol and owns the output-token efficiency curve. Meta's Muse Spark 1.3 is the first time the Llama people have looked like a closed-frontier lab that also happens to leak gravity into the open ecosystem. Google is iterating Flash the way a supermarket iterates milk: often, cheap, everywhere.
The price of a point
DeepSeek V4 Pro still does the rude thing on the cost chart. Closed models near the top cost twenty to forty-five times more per task. That gap is the entire open-weight business model, and it has not closed. What has closed is the quality gap that used to let frontier labs pretend open weights were a hobby. Kimi K3 walked onto the 50-line in July. GLM-5.3 sits on the Pareto front at a couple of dollars per million tokens and four hundred tokens a second.
The lead used to last a year. This summer it lasted six weeks.
The editorial problem is the half-life of being first. The news is not a model. We will reprint the table every issue until the table stops moving.