News

Market & Business

CONFIRMED: Claude Opus 5.5 leads new Intelligence Index v5 from Artificial Analysis

Artificial Analysis launched Intelligence Index v5, the biggest overhaul of its AI benchmark since January, now built 60% on private test sets. Claude Opus 5.5 leads at 58 points, and the gap between open and closed models quadrupled.

Ilustração abstrata de barras de vidro translúcidas em tons de azul representando um ranking de inteligência artificial

The scoreboard changed, and that is good news for buyers

Artificial Analysis, one of the few independent organizations the entire AI market relies on to compare models, just changed the measuring stick. On Tuesday it launched Intelligence Index v5, its biggest upgrade since version 4 in January, and the part that matters outside the benchmark bubble is simple: it got much harder for a lab to train a model specifically to ace the test.

A closed exam instead of a memorized one

Until now, most AI benchmarks were public: any lab could see the questions, suspect they were sitting in its training data and tune the model for that specific set, inflating the score without a real capability gain. In v5, 60 percent of the index now runs on private datasets or held out answers that labs never see before the evaluation. Artificial Analysis also swapped in new criteria: FrontierCode 2.0, built by Cognition for end to end real world coding, and Terminal Bench Science, focused on scientific tasks executed in a terminal.

Who is on top now

Under the new ruler, Anthropic's Claude Opus 5.5, running at max effort, leads with 58 points, ahead of OpenAI's GPT-6 Astra, also at max effort, with 57, and GPT-6.1 Sol with 55. Anthropic and OpenAI hold the top five spots in the ranking. Google's Gemini 4 Argon, in high mode, follows with 50 points. The gap between first and second place is small but symbolic: the race between the two labs keeps moving point by point even under a much stricter evaluation standard.

The gap between open and closed models doubled

The number that stands out most for anyone tracking open source AI is different: the distance between the best open weight model and the best closed model jumped from 3 points in v4 to 12 points in v5. Moonshot's Kimi K3.1 Pro leads among open models with 46 points, but its weights have not been published yet. Among open models already available for download, DeepSeek V4.1 Flash and GLM-5.3 tie at 32 points, 26 points behind the overall leader. In other words, the claim that open source has already caught up to closed models loses ground exactly when the ruler got harder to game.

The hidden price of climbing the ranking

Artificial Analysis also exposed how much a high score costs. Claude Opus 5.5 ranges from 35 points at low effort to 58 at max effort, a performance jump that comes with roughly ten times higher inference cost. Median output tokens per task rose from 35 thousand to 59 thousand between versions, growth driven mostly by agentic evaluations: they carry 55 percent of the final score but account for about 90 percent of the tokens generated in testing. In practice, climbing the ranking increasingly means letting the model think and act longer, not just a jump in architecture.

What this changes for anyone choosing a model

For whoever decides which AI goes into production, the practical reading is twofold. First, the ranking got more trustworthy to bet money on, because it cuts the chance a model only looks good because it saw the exam beforehand. Second, comparing scores without comparing the cost of the effort mode used has become a trap: a model can lead at max effort and lose its financial edge the moment effort gets dialed down to fit a budget. Artificial Analysis also launched the Progress Index v1 in parallel, comparing capability of models released since November 2022 on a single scale, useful for tracking historical evolution, though the Intelligence Index remains the reference for immediate usage decisions.

Sources

Artificial Analysis: Announcing Artificial Analysis Intelligence Index v5; Artificial Analysis on X: Claude Opus 5.5 takes the top spot; TestingCatalog on X: Daily AI Brief, October 28.