Artificial Analysis published its Intelligence Index v5 on Oct. 272. The ranking crowned a new leader for the first time since January. Claude Opus 5.5 topped the board, with GPT-6 Astra one step behind1.
Anthropic and OpenAI now hold every place among the top five1. Most of the exam stays hidden from view, six of every ten questions sit behind a wall only Artificial Analysis can open. The firm built that wall on purpose. It wanted to stop labs from training on the very questions meant to test them3.
A Score Built Mostly on Secret Tests
Public tests grow old fast. Once their answers circulate online, a model can memorize the exam and skip the skill underneath. Its score still climbs, even as the ability behind it stays put. That is the problem.
Artificial Analysis calls that problem contamination3. Version five answers it with private test sets and held-out answers, now worth 60 percent of the final score1. Two tests joined the index for the first time this round.
FrontierCode 2.0 grades an agent on a real open-source repository, and a human maintainer decides whether the fix would get merged1. Terminal-Bench-Science sends an agent through a run of science tasks inside a plain terminal, graded step by step.
Artificial Analysis@ArtificialAnlysAnnouncing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets
The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks… pic.twitter.com/oA4tsXQOC6
One upgrade retired a well-known exam. Humanity's Last Exam gave way to HLE-Diamond, a thousand-question cut split evenly between reasoning and knowledge1.
Dan Hendrycks, executive and research director at the Center for AI Safety, helped build the original test. The idea began with a talk with Elon Musk, who found older benchmarks too easy67. Scale AI and Hendrycks's team pulled its questions from experts across more than fifty countries.
The Points Every Open Model Lost
These new rules touched every model differently. Before the rebuild, the gap between the top model and the best open model spanned three points. Under version five, it spans 121.
Kimi K3.1 Pro, built by Moonshot AI, still leads every open model at 46 points1. That mark sits nine points lower than before the rebuild, and its weights have yet to ship5. DeepSeek V4.1 Flash fell 13 points to 32, tied there with GLM-5.3 among the models whose weights do ship. Both now sit 26 points behind Claude Opus 5.5.
Jakob Steinschaden, editor-in-chief of the Vienna outlet Trending Topics, tallied the slide elsewhere on the board5. Xiaomi's MiMo-V2.6-Pro lost 16 points and settled at 30. Meta's Muse Spark Open dropped 14 points to 25.
Four Open Models Lost 9 to 16 Points When Artificial Analysis Rebuilt Its Index
Intelligence Index score before and after v5, four open-weight models, October 2026
Each row is one model: the blue dot is its score before v5, the magenta dot is its score after, and the rod is the points it lost. Higher is better.
Artificial Analysis rescored every model under the new rules at once, so these are the same models tested twice under one shared standard.
Sources: Artificial Analysis (X/Twitter) [1]; Trending Topics [5]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Before v5 | After v5 |
|---|---|---|
| Kimi K3.1 Pro | 55 pts | 46 pts |
| DeepSeek V4.1 Flash | 45 pts | 32 pts |
| MiMo-V2.6-Pro | 46 pts | 30 pts |
| Muse Spark Open | 39 pts | 25 pts |
Who Benefits When the Test Stays Secret
A private test blocks cheating. It also blocks checking, because only the firm that wrote the questions can see them. A buyer who wants to confirm last month's score has one option. That option means trusting the same company that built the test and sells access to it.
That is a lot to ask.
Nitish Garg, founder and chief executive of the AI staffing platform CellCog, backed that trade. "A model cannot be trained on questions it has never seen," Garg wrote of the new index. "With 60% of v5 on private sets or held-out answers, a high score is harder to earn by memorizing public benchmarks"4.
Steinschaden read the same shift and looked past the labs that gained. "And because most of the test data is now secret, the results can no longer be independently verified from the outside," he wrote5. His own tally shows why that worry lands hardest on open models. Four of them lost nine points or more apiece, while GPT-6 Astra and GPT-6.1 Sol each climbed.
A model that trails badly overall can still win a share of the board2. Claude Opus 5.5 leads FrontierCode 2.0 at 46 percent and HLE-Diamond at 68 percent. GPT-6 Astra leads Terminal-Bench-Science at 63 percent.
Kimi K3.1 Pro leads Terminal-Bench 4.0 at 67 percent2. It is the one major test in version five that anyone outside Artificial Analysis can still read. That test survived the rebuild. Most of its neighbors stayed behind the wall.
Kimi K3.1 Pro Leads One of Four Tests While Trailing by 26 Points Overall
Top score on four of Intelligence Index v5's nine evaluations, percent of tasks solved, October 2026
Each bar is the single highest score on that test; the magenta bar is the only one an open-weight model leads, and the blue bars are each led by a closed model. Higher is better.
Each bar is a different test with its own task set, so the four scores each measure a separate skill on its own scale.
Source: Artificial Analysis [2]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| HLE-Diamond (Opus 5.5) | 68% |
| Terminal-Bench 4.0 (Kimi) | 67% |
| Term.-Bench-Science (Astra) | 63% |
| FrontierCode 2.0 (Opus 5.5) | 46% |
What to Do
Read the breakdown behind the composite score before acting on it. A buyer choosing a coding agent should weigh FrontierCode 2.0 and Terminal-Bench 4.0 on their own terms. Claude Opus 5.5, GPT-6 Astra and Kimi K3.1 Pro each lead a different piece of this index.
Give an open-weight vendor a fair hearing before a 12-point gap rules it out. Ask how much of that gap sits inside tests you can read for yourself. The rest may sit behind a wall only one firm can open.
By the Numbers
- Claude Opus 5.5 scored 58 points, the new top mark on Intelligence Index v51.
- GPT-6 Astra trailed by one point, finishing at 571.
- Private test sets and held-out answers now supply 60 percent of the score1.
- The gap between the leader and the best open model widened from three points to 121.
- DeepSeek V4.1 Flash and GLM-5.3 trail the leader by 26 points among open models whose weights ship1.
- Kimi K3.1 Pro's 67 percent on Terminal-Bench 4.0 is the top mark of any model on the index's one surviving public test2.
- FrontierCode 2.0, the new private coding benchmark, draws its tasks from 132 real repositories1.
- Scale AI and the Center for AI Safety wrote 2,500 questions into the original Humanity's Last Exam, later cut down for HLE-Diamond6.
What to Watch
Watch for a second benchmarking group to run its own private tests on these same models. A matching order would answer Steinschaden's complaint better than any statement from Artificial Analysis itself could.
Moonshot AI and DeepSeek could also publish their own results on FrontierCode 2.0 or Terminal-Bench-Science, the two tests closed to outside review. A published score from either lab would show whether the open-weight slide is a measure of skill or a measure of access. Either answer would tell a buyer more than one more press release ever could. That is the real test.
Sources
- Artificial Analysis, "Announcing Artificial Analysis Intelligence Index v5," Artificial Analysis (X/Twitter), Oct. 28, 2026, https://x.com/ArtificialAnlys/status/2115265133467635877
- Artificial Analysis, "Announcing Artificial Analysis Intelligence Index v5," Artificial Analysis, Oct. 27, 2026, https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5
- Artificial Analysis, "Artificial Analysis Intelligence Benchmarking Methodology," Artificial Analysis, Oct. 27, 2026, https://artificialanalysis.ai/methodology/intelligence-benchmarking
- Nitish Garg, "Intelligence Index v5: Claude Opus 5.5 Leads at 58," CellCog, Oct. 28, 2026, https://cellcog.ai/blog/artificial-analysis-intelligence-index-v5/
- Jakob Steinschaden, "Artificial Analysis's New A.I. Index Penalizes Open-Weight Models," Trending Topics, Oct. 28, 2026, https://www.trendingtopics.eu/artificial-analysis-index-v5-open-weight-models/
- Center for AI Safety and Scale AI, "Humanity's Last Exam," lastexam.ai, April 3, 2025, https://lastexam.ai/
- Center for AI Safety, "About Us," Center for AI Safety, https://safe.ai/about
- ai.engineer, "Micah Hill-Smith," ai.engineer, https://ai.engineer/speakers/micah-hill-smith