The Artificial Analysis Intelligence Index v5, released on October 27, 2026, is a new version of the independent composite score that ranks language models. It averages nine evaluations split evenly across four groups (Knowledge Work, Coding, Science and General, 25% each), adds FrontierCode 2.0 and Terminal-Bench-Science, swaps Humanity's Last Exam for HLE-Diamond, and drops GDPval-AA, SciCode and AA-LCR. For the first time, 60% of the index rests on private test sets or held out answers. Scores are lower and spread wider than in v4.3, so a v5 number cannot be compared with any older one.
At launch, Claude Opus 5.5 (max) led at 58, ahead of GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55 (launch post). When we checked the live leaderboard on the morning of October 29, a newly added Gemini 4 Carbon (high) sat on top at 59. This guide explains what is inside v5, what changed from v4.3.2, and how to read the number without fooling yourself. For how Artificial Analysis compares with other boards, see our LLM leaderboards compared guide.
What is the Artificial Analysis Intelligence Index v5?
Artificial Analysis runs every model itself, under the same prompts and settings, and combines the results into one weighted average. The methodology page describes the index as mainly text based and English language. Models that accept images do get images in some tests, such as the HLE-Diamond image questions and the GDP.pdf page images. Multilingual, vision and speech results are published separately and do not count.
Reasoning models run at temperature 0.6 and other models at 0, unless a lab recommends another setting. Most tests use pass@1, averaged over repeats where a test is run more than once. Agentic tests run in e2b sandboxes, mostly inside Artificial Analysis's own open source harness, Stirrup.
Each effort setting of a model appears as its own row. "Claude Opus 5.5 (max)" and "Claude Opus 5.5 (low)" are scored separately, and in v5 that gap is large, which we come back to below.
The nine evaluations and their weights
| Group (25% each) | Evaluation | Weight | What it tests | Private |
|---|---|---|---|---|
| Knowledge Work | AA-Briefcase v1.1 | 15% | 91 multiweek business tasks across 4 scenarios, graded into an Elo score | Yes |
| Knowledge Work | AutomationBench-AA | 5% | 657 SaaS workflow tasks over REST APIs; breaking a guardrail scores zero | Yes |
| Knowledge Work | GDP.pdf | 5% | 100 tasks grounded in long PDFs, all criteria must pass | No |
| Coding | FrontierCode 2.0 | 15% | 132 changes to real open source repositories, graded on mergeability | Yes |
| Coding | Terminal-Bench 4.0 | 10% | 66 terminal tasks, every test must pass | No |
| Science | Terminal-Bench-Science 0.1 | 15% | 70 research workflow tasks in five sciences | No |
| Science | CritPt v2 (hard subset) | 10% | 40 research level physics problems | Yes |
| General | AA-Omniscience v1.1 | 15% | 3,000 knowledge questions; 10% accuracy, 5% non hallucination | Yes |
| General | HLE-Diamond | 10% | 1,000 expert questions, no tools | No |
Weights and task counts come from the methodology page as of October 29, 2026.
The two new benchmarks
FrontierCode 2.0 is Cognition's private coding benchmark, launched the same day as v5 (Cognition's post). Maintainers of open source projects wrote the tasks and the grading criteria, and the question is whether they would merge the agent's change. Compared with version 1.1, Cognition says the reference solutions touch 67% more files and 89% more lines of code, in codebases 25% larger, with 43% more blocking criteria. Artificial Analysis runs all 132 tasks three times in Stirrup with GPT-6 Luna (medium) as the judge. Any attempt that fails a blocking check, or that its trajectory scanner flags for looking up the solution, scores zero. Our FrontierCode benchmark guide explains the grading in detail, and the FrontierCode tool page tracks the official board.
Terminal-Bench-Science 0.1 comes from Stanford researchers with the Terminal-Bench and Harbor team (project site). Its 70 tasks are drawn from real research work: 19 in life sciences, 17 in physical sciences, 17 in mathematical sciences, 9 in engineering and 8 in earth sciences. A task passes only if every one of its tests passes, and agents get up to 1,000 steps with a 30 second limit on each shell command.
The three upgrades
HLE-Diamond replaces Humanity's Last Exam. It is the refined 1,000 question version from the Center for AI Safety and Scale AI, split into 500 reasoning and 500 knowledge questions, with 272 image questions (announcement). The old index used 2,158 text only HLE questions, so the two scores are not comparable. Our HLE Diamond guide covers the dataset and the official board.
AA-Omniscience v1.1 shrinks from 6,000 to 3,000 questions across 42 topics. Artificial Analysis removed saturated questions and the 600 questions it had published in the v1 public set, added new ones and fixed ambiguous ones. A correct answer earns points, a confident wrong answer loses points, and saying "I don't know" is neutral. A refreshed public sample is on Hugging Face. If hallucination scoring is new to you, our LLM hallucination detection guide gives the background.
CritPt v2 (hard subset) keeps 40 of the 70 CritPt physics problems. The 30 removed were the ones seven frontier models solved in all five attempts when CritPt v2 launched. Models work in an offline Python sandbox.
What changed from v4.3.2
The version history lists the changes. In short:
| Item | v4.3.2 (Sept to Oct 2026) | v5.0 (Oct 27, 2026) |
|---|---|---|
| Groups | Agents 30%, Coding 20%, Scientific Reasoning 20%, General 30% | Knowledge Work, Coding, Science, General at 25% each |
| Evaluations | 10 | 9 |
| Coding | Terminal-Bench 4.0, SciCode | FrontierCode 2.0, Terminal-Bench 4.0 |
| Science | Humanity's Last Exam, CritPt | Terminal-Bench-Science, CritPt v2 hard subset |
| Knowledge work | GDPval-AA v2.1, AA-Briefcase, AutomationBench-AA | AA-Briefcase, AutomationBench-AA, GDP.pdf (cut from 10% to 5%) |
| General | AA-Omniscience, GDP.pdf, AA-LCR | AA-Omniscience v1.1, HLE-Diamond |
Removed from the index: GDPval-AA v2.1, SciCode, AA-LCR v1.1 and the original Humanity's Last Exam. GDPval-AA still runs as a standalone evaluation, so its leaderboard keeps updating; our GDPval benchmark guide explains how it works. The Agents group is now called Knowledge Work, and Scientific Reasoning is now Science.
The private share is the headline change. Private sets are harder to train on by accident or on purpose, which is the main worry with public benchmarks. The tradeoff is that you cannot inspect most of the questions yourself, so you are trusting the people who run the tests.
Key results at launch
From the October 27 launch post:
- Top of the board: Claude Opus 5.5 (max) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55. Anthropic and OpenAI held the top five spots, and Gemini 4 Argon (high) was next at 50.
- No model leads everything. Claude Opus 5.5 led FrontierCode 2.0 (46%), HLE-Diamond (68%) and AA-Omniscience v1.1. GPT-6 Astra led Terminal-Bench-Science (63%) and the CritPt v2 hard subset (72%). Claude Sonnet 5.5 led AA-Briefcase, Kimi K3.1 Pro led Terminal-Bench 4.0 (67%), Gemini 4 Argon led AutomationBench-AA (78%) and Muse Wonder Full led GDP.pdf (33% all pass).
- Open weights fell further behind. The gap to the leader grew from 3 points in v4.3 to 12. Kimi K3.1 Pro, whose weights were not yet released, scored 46. Among models with released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) led at 32, below Claude Haiku 5.5 (max) at 36.
- Cost per task varies widely. Claude Opus 5.5 (max) cost $10.04 per index task, GPT-6 Astra (max) $6.24 and GPT-6.1 Sol (max) $1.27. Among the ten highest scoring models, GPT-6.1 Terra (max) was cheapest at $0.96 and Claude Fable 5.1 (max) the most expensive at $13.04.
Rankings move as new models are added. On October 29 the live board listed Gemini 4 Carbon (high) at 59, then Claude Opus 5.5 (max with fallback) at 58 and GPT-6 Astra (max) at 57. Always check the date on any ranking you cite. Model by model comparisons are outside this guide; we focus on what the score measures.
Why effort settings matter more in v5
The same model can look like two different products depending on its effort setting. Artificial Analysis reports that Claude Opus 5.5 scores 35 at low effort and 58 at max, at about ten times the cost per task ($0.98 against $10.04). Claude Sonnet 5.5 now spans 29 points from low to max, up from 20 in v4.3, and GPT-6.1 Sol's spread doubled from 10 to 20 points.
The tasks also demand more work. The median model now writes about 59,000 output tokens per index task, up from 35,000 in v4.3. The four agentic evaluations (Terminal-Bench-Science, Terminal-Bench 4.0, AA-Briefcase and FrontierCode 2.0) carry 55% of the weight but produce about 90% of output tokens, and Terminal-Bench-Science alone accounts for about a third.
So a headline score only describes a model at one effort setting. If you plan to run a model at medium effort to save money, compare the medium rows, not the max ones.
The new Progress Index
Artificial Analysis also launched the Progress Index v1 alongside v5 (methodology). It exists because the Intelligence Index changes its test set with each version, so its scores are not meant for long run comparison.
The Progress Index uses item response theory. It estimates each model's capability and each task's difficulty together, linking models through the tasks they share, so a 2023 model and a 2026 model can sit on one scale without taking the same tests. The scale is anchored so that o3-mini (high) scores 220 and GPT-5.5 (xhigh) scores 320. A benchmark needs results from at least four models, and a model needs results on at least four benchmarks, to be included. Artificial Analysis credits learnings from Epoch AI's Epoch Capabilities Index; see the Epoch AI Benchmarking Hub for that work.
Artificial Analysis's own advice: use the Progress Index to study trends over time, and the Intelligence Index to pick a model today.
How to read a v5 score
- Cite the version. Write "Intelligence Index v5" next to any number. A model that scored in the high 50s on v4.3 and scores 58 on v5 did not stand still or improve; the test changed.
- Look at the components that match your work. A coding team should care about FrontierCode 2.0 and Terminal-Bench 4.0 more than about CritPt. A research assistant use case leans on AA-Omniscience and HLE-Diamond. Our coding agent benchmark rankings collect the coding side in one table.
- Check the effort setting and cost per task together. A three point lead can cost eight times as much per task.
- Treat small gaps as ties. One or two points between models is within the range where run to run noise and judge choices matter. Several components use LLM judges, which have known biases; see our LLM as a judge guide.
- Run your own tests before you commit. A public index tells you who to shortlist, not who will win on your data. Our guide on how to evaluate LLM model performance shows how to build a small private test set.
Limits to keep in mind
- You cannot audit most of it. With 60% of the weight on private sets, outsiders can read the methods but not the questions.
- One lab runs everything. That keeps conditions identical, but it also means one set of harness choices (Stirrup, mini-swe-agent, turn caps) shapes every score. A model tuned for a different agent setup may score lower here than in its own tools.
- English and text first. The index does not tell you how a model handles other languages, speech or most vision work.
- The test will change again. v4 lasted from January to October 2026 with several point releases. Expect v5 to be revised as models saturate parts of it.
If you want a second opinion from a board built differently, our LiveBench vs Artificial Analysis guide compares the two, and LLM benchmarks explained covers what individual scores mean.
FAQ
When did Artificial Analysis release Intelligence Index v5?
Artificial Analysis announced Intelligence Index v5 on October 27, 2026, the same day Cognition launched FrontierCode 2.0. It replaced v4.3.2, which had been in use since September 2026.
Can I compare a v5 score with a v4 score?
No. Each version uses a different set of tests and weights, and Artificial Analysis says scores across versions are not intended for long run comparison. Use the new Progress Index if you want to compare models over time.
Why did GDPval-AA leave the Intelligence Index?
Artificial Analysis lists GDPval-AA v2.1 as removed in v5.0 without giving a detailed reason in the methodology. It still runs GDPval-AA as a standalone evaluation, and knowledge work in the index is now covered by AA-Briefcase, AutomationBench-AA and GDP.pdf.
Which model leads the Intelligence Index v5?
At launch on October 27, Claude Opus 5.5 (max) led at 58. On October 29 the live leaderboard showed a newly added Gemini 4 Carbon (high) at 59, just ahead of Claude Opus 5.5. The order changes as models are added, so check the leaderboard date.
What is the difference between the Intelligence Index and the Progress Index?
The Intelligence Index measures today's frontier on a test set that changes with each version. The Progress Index v1 uses item response theory to put models released since November 2022 on one stable scale, so it suits trend analysis rather than buying decisions.
Is the Artificial Analysis Intelligence Index free to view?
Yes. The leaderboard, the methodology and the launch posts are public on artificialanalysis.ai. Most v5 test questions are private, but Artificial Analysis publishes sample datasets such as the AA-Omniscience public set on Hugging Face.
