Claude Opus 4.8

Six weeks after Opus 4.7, Anthropic shipped Opus 4.8 and, at launch, edged past OpenAI to the top of the Artificial Analysis Intelligence Index — the one composite score run by a disinterested third party — at 61.4, about a point clear of GPT-5.5. It is the first time the independent crown has sat with Anthropic rather than OpenAI.

Anthropic's headline number isn't the leaderboard win — it's a model ~4x less likely than 4.7 to let a flaw it wrote pass unflagged, turning that flag into a signal for human review.

The crown is the thinner half of the story. The margin is roughly a point, the index moves constantly, and a newer Claude variant had already passed 4.8 within days. The larger coding gaps Anthropic quotes — a wide lead on the hard SWE-Bench Pro test, being the only model to finish every task in its own agent benchmark — are the company's own numbers, not anyone else's, and several analysts note such cross-vendor gaps shrink once you account for how each model was run.

What Anthropic actually led with is quieter: calibration. The headline feature is that the model is roughly four times less likely than its predecessor to let a flaw in code it wrote pass without flagging it. That is not better code — it is a model that tells you where it is unsure. One analyst's read: a flagged output becomes a first-class signal for human review rather than something to catch later. The day it could have led with its first independent #1, the lab instead opened with a model that volunteers its own doubt.