A Guide to Artificial Analysis Intelligence Index v5: What It Measures and Why It Matters

Artificial Analysis launched Intelligence Index v5 today, its biggest overhaul since January. For the first time, most of the Index runs on private data, and the results show frontier models producing far more tokens per task than nine months ago. Both shifts land directly on infrastructure.

At GMI Cloud, we serve inference for teams that choose models on evidence. Independent, repeatable measurement helps our customers pick well and helps us plan the infrastructure those picks depend on. Congratulations to the Artificial Analysis team on a release that makes measurement harder to game and closer to real work.

What changed in v5

The headline: 60% of the Index now runs on private test sets or held-out answers. That is the first time private data makes up the majority, and it directly addresses contamination as models train on more of the public web.

The Index now rests on four equal pillars of 25% each, built from nine evaluations:

Two design choices stand out to us. FrontierCode 2.0 scores a task only when a repository maintainer would merge the change. AA-Omniscience v1.1 rewards a model for admitting uncertainty as well as for answering correctly. Both measure what customers feel in production.

What the results tell us

Leadership is split: across the nine evaluations, six different models hold at least one top spot.

Overall, Claude Opus 5.5 (max) leads at 58, with GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55.

Intelligence Index v5 scores for selected models, with published cost per Index task. Data: Artificial Analysis.

Cost matters as much as rank. Among the 10 highest-scoring models, cost per Index task ranges from $0.96 to $13.04, a spread of more than 13x. GPT-6.1 Sol (max) lands within 4 points of the leader at about one eighth of the cost.

For our customers, the lesson is simple: the best model depends on the workload. A coding agent, a document pipeline and a research agent may each want a different model, and often more than one.

Why this matters for infrastructure

v5 makes one trend clear: frontier work is getting longer, more agentic and more token-hungry.

More tokens per task. Every model measured produces more output tokens per Index task than in v4.3. The median rose by about two thirds, from 35k to 59k.

Agents drive the load. The four agentic evaluations carry 55% of the Index weight and generate a median of about 90% of output tokens.

Effort settings change outcomes. Claude Opus 5.5 scores 35 at low effort and 58 at max, at roughly 10x the cost per task ($0.98 versus $10.04). How much compute a team is willing to spend now shapes the result it gets.

All three point the same way. Teams building agents need inference that holds up through long, multi-step runs, with throughput they can count on and costs they can predict.

How GMI Cloud is building for it

As an NVIDIA Exemplar Cloud, we run our own GPU infrastructure across the US and APAC. That lets us give teams the two things v5 says they need: room to run long agentic workloads, and a clear view of what each run costs.

Prime Inference is our next step for teams running long agentic workloads. More to share soon.

Model choice through MaaS and Model Router lets teams match each workload to the model that fits it and switch as results change.

AgentBox is designed to run multi-session agents that carry work across sessions, beyond single prompts. That is exactly what v5's Knowledge work pillar (25%) now tests: AA-Briefcase alone measures multi-week professional projects, and it carries 15% of the Index.

Benchmarks like Intelligence Index v5 keep the whole industry honest, and v5 confirms the direction we have been building toward: longer runs, more agents, more tokens per task. Explore the full results at artificialanalysis.ai, and write to [email protected] when you want to run those workloads on GMI Cloud.