Artificial Analysis has changed the tests behind its Intelligence Index. For teams choosing an AI model, the immediate consequence is to compare results within the same version and then check the tasks, effort setting and cost that produced them.
The organization's v5 announcement is dated October 27, 2026. Its discovery-feed timestamp is October 28. The page's own date establishes the announcement; the later feed entry does not make this a second release.
What the composite now rewards
The updated index assigns equal weight to knowledge work, coding, science and general capabilities. Artificial Analysis says 60% of its weight now comes from evaluations using private test sets or held-out answers. That is a property of the evaluation design, not a guarantee that every contamination risk has been removed.
Two additions illustrate the shift. FrontierCode 2.0 examines repository changes, while Terminal-Bench-Science tests scientific tasks in a terminal. Cognition describes FrontierCode's goal in terms of whether a maintainer would accept an agent's contribution, including design and integration judgments.
For a software team, that is a reason to inspect what a coding subscore measures. It does not establish that a high overall score predicts acceptance in its own repository. A composite spreads attention across several kinds of work, while a deployment may depend heavily on one.
Compare configurations before costs
The published methodology sets out the component weights and testing arrangements. The launch analysis also shows that reasoning effort changes both scores and resource use, and reports more output tokens per task than under v4.3.
A useful comparison therefore records the model and effort level together. A cost-per-token figure alone cannot tell a team how much it will spend to complete its task. Tool use, repeated attempts and human review belong in that local evaluation too.
Use v5 to form a shortlist, then run representative work with explicit acceptance criteria. Keep track of successful outputs and the work needed to correct failures. This is an evaluation proposal, not a claim that TechKili has reproduced the benchmark.
A separate scale for progress over time
Artificial Analysis's Progress Index methodology explicitly says Intelligence Index scores across versions are not intended for long-run comparison. The Progress Index instead estimates model capability and task difficulty on a shared scale.
That distinction answers two different reader needs: compare current candidates with the current Intelligence Index, and examine longer-term capability trends with the Progress Index. Neither measure alone establishes a business return; our analysis of AI productivity and financial outcomes explains why those require their own evidence.
Sources
Related Articles
Anthropic's IPO filing ties much of its $80.7 billion loss to valuation changes
The public S-1 reports a much smaller operating loss and positive operating cash flow for the same six months. These measures explain different parts of the business.
Apollo's AI payoff claim hinges on how productivity is measured
Torsten Slok's October 24 note questions broad AI returns. Its index comparisons and productivity charts are useful context, but they do not measure a particular company's deployment.
Google Personal Intelligence can carry connected-app context into ads
Google's October policy and Search help allow derived context from connected apps to inform ads. Reviewing connections, ad settings and saved history involves separate choices.
Etsy bars AI-written functional guides and asks listings to show the real result
Etsy's October policy update separates permitted AI-assisted art from functional instructions, while tightening the evidence sellers must show in product images.