The Artificial Analysis Intelligence Index v5 is a reset of one of the most-quoted AI leaderboards, and it changes how you should read any “best model” claim. Announced on October 27, 2026, it adds two new tests, upgrades three more, and builds most of the score on private questions.
The short version: scores from v5 can’t be compared with older versions, and Google’s Gemini 4 Carbon now leads by just one point.
Quick facts, based on the Artificial Analysis changelog:
- Announced: October 27, 2026
- Changes: 2 new benchmarks, 3 upgraded ones, 3 removed
- Most of the index now uses private datasets (no exact share given)
- Top score: Gemini 4 Carbon (High) at 59
Those four lines tell you what moved. The sections below explain why it matters when you pick an AI model.
What the Intelligence Index v5 actually changed
The index now combines nine benchmarks. Two are new: FrontierCode 2.0 and Terminal-Bench-Science.
FrontierCode 2.0 is Cognition’s private test of real-world coding tasks, where an agent’s change counts only if a repository maintainer would merge it. Terminal-Bench-Science has 70 agentic research tasks run in a terminal.
Three existing tests were upgraded. HLE-Diamond replaces Humanity’s Last Exam with a refined 1,000-question set. AA-Omniscience got a new dataset, and CritPt moved to a private 40-question hard subset.
Why private questions matter for AI rankings
Public benchmark questions leak into training data over time. A model that has seen the answers looks smarter than it is.
The changelog says the update removed saturated and public-set questions from AA-Omniscience. It also says most of the index is built on private datasets. The page doesn’t explain its reasoning at length, so treat the motive as an inference: fewer public questions means fewer ways to game the score.
Who leads the new rankings
According to XenoSpectrum, Gemini 4 Carbon (High) scores 59, one point ahead of Claude Opus 5.5 (Max) at 58. Here is how the top of the board looks:
| Model | v5 score | Note |
|---|---|---|
| Gemini 4 Carbon (High) | 59 | First place |
| Claude Opus 5.5 (Max) | 58 | One point behind |
| Claude Opus 5.5 (Xhigh) | 57 | Same model, lower setting |
| GPT-6 Astra | 57 | Tied with Opus Xhigh |
| Kimi K3.1 Pro | 46 | Per the changelog |
Notice that Opus 5.5 appears twice. Reasoning settings change scores, so one model can sit in two places on the same chart.
We covered the launch in our Gemini 4 Carbon release date guide, and the open-weight side in our Kimi K3.1 Pro benchmarks post.
Why a one-point lead is not a clear win
Carbon is first on four of the nine tests, including the largest lead on AA-Omniscience. It is second on Terminal-Bench 4.0, and it isn’t first on FrontierCode 2.0, CritPt, AA-Briefcase or GDP.pdf, per XenoSpectrum.
So a single average hides a lot. If you mostly want coding help, FrontierCode 2.0 matters more to you than the headline number.
Cost is the other half. XenoSpectrum lists per-task cost at list prices: GPT-6 Astra (Max) $6.25, Gemini 4 Carbon $7.76 and Claude Opus 5.5 (Max) $10.04. A one-point gap can be smaller than a $2 to $4 cost gap.
Frequently Asked Questions
What is the Artificial Analysis Intelligence Index?
It’s a combined score from several AI benchmarks, used to compare language models. Version 5 uses nine tests.
Can I compare v5 scores with older scores?
No. v5 adds and changes questions, so its totals aren’t directly comparable with the previous version. Carbon’s 59 versus Gemini 4 Argon’s 50 is a new-index comparison.
Which AI model is first on the Intelligence Index v5?
Gemini 4 Carbon (High) at 59, one point ahead of Claude Opus 5.5 (Max).
Why does the index use private datasets?
Private questions can’t be memorised from the public web. The changelog doesn’t spell out the reason, but it fits that goal.
Verdict: use the index to shortlist models, not to crown one. A one-point lead, cost per task and your own workload should decide, so run your real prompts on the top two or three before you commit.
Leave a Reply