Artificial Analysis Intelligence Index v5: Opus 5.5 leads at 58
The leaderboard that settles most arguments about frontier models just changed what it measures — and, below the top spot, the order changed with it. Artificial Analysis shipped Intelligence Index v5 on Tuesday, its biggest rewrite since v4 arrived in January and the first version where most of the score rests on test sets nobody can train against.
Claude Opus 5.5 (max) holds first at 58, but GPT-6 Astra (max) is now second at 57 and GPT-6.1 Sol (max) third at 55, according to the announcement. Anthropic and OpenAI take all five top slots, with Gemini 4 Argon (high) the next model at 50. The leader survived the rewrite; most of the board underneath it did not.
What actually changed. Two benchmarks came in: FrontierCode 2.0, Cognition's private set of 132 real-world coding tasks graded on whether a repository maintainer would merge the patch, and Terminal-Bench-Science, 70 expert-authored research workflows run in a terminal. Three were upgraded. Humanity's Last Exam gave way to HLE-Diamond, a refined 1,000-question version; AA-Omniscience moved to v1.1 with saturated and leaked questions stripped out; CritPt v2 contributed a private 40-question hard subset after the original 70 turned out to be solvable. Three benchmarks left: GDPval-AA v2.1, SciCode and AA-LCR.
The weighting is now four equal 25% quarters — knowledge work, coding, science, general — which means coding and science carry far more of the score than they did in v4. The headline number in the methodology is that 60% of the index now sits on private test sets or held-out answers. That is the whole point of the rewrite: every public benchmark eventually gets absorbed into training data, and a leaderboard that cannot be contaminated is a different product from one that can.
The reshuffle is the story. In v4.3 the top five read Opus 5.5 at 58, Claude Sonnet 5.5 and Meta's Muse Wonder Full at 56, Kimi K3.1 Pro at 55 and Claude Fable 5.1 at 53 — as laid out in a side-by-side of the old and new tables. Under v5, Astra jumps to second, Sol into third, and Sonnet 5.5, Muse Wonder Full, Kimi and Fable all fall out of the top five. A model that could not beat Fable 5.1 in October now trails the leader by a single point. Either the new suite measures something the old one missed, or the old one was flattering the wrong labs. The reordering is large enough that anyone citing a pre-v5 score this week is citing a different index.
Open weights lose ground, badly. The best open-weights model, Kimi K3.1 Pro, scores 46 — twelve points behind Opus 5.5, against a three-point gap in v4.3. Among models whose weights are actually downloadable, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) top out at 32, which is below Claude Haiku 5.5 (max, 36). We wrote three weeks ago that Kimi had pulled within three points of the frontier — Kimi K3.1 Pro: 55 on the AI index, 3 points behind Opus 5.5 — and the honest follow-up is that the gap was partly an artifact of which tests were being counted. The new benchmarks are agentic and scientific, and that is where open models are furthest behind: Kimi scores 37% on HLE-Diamond against 68% for Opus 5.5, and 34% on Terminal-Bench-Science against 63% for Astra.
Cost is where the top of the table gets interesting. Per-task cost among the ten highest-scoring models spans more than 13x, from $0.96 for GPT-6.1 Terra (max) to $13.04 for Claude Fable 5.1 (max). Opus 5.5 costs $10.04 per index task, Astra $6.24, Sol $1.27. Sol sits four points behind the leader at roughly an eighth of the cost per task, which is the number procurement teams should be reading rather than the ranking.
Effort settings now swing the score more than they used to. Opus 5.5 scores 35 at low effort and 58 at max, and the cost per task runs from $0.98 to $10.04 across that range — roughly 10x for 23 points. Sonnet 5.5 spans 29 points from low to max, up from 20 in v4.3, and Sol's spread doubles from 10 to 20 points. Anyone comparing models on this index has to state the effort setting, or the comparison is meaningless; the same model can look mid-table or frontier depending on a dropdown.
v5 tasks are also far heavier. Median output tokens per task rise two thirds, from 35,000 to 59,000, because the agentic evaluations now carry 55% of the index weight while consuming about 90% of the tokens. Terminal-Bench-Science alone accounts for roughly a third of output tokens across the suite. That is a quiet repricing of what "running a benchmark" costs a lab or an independent evaluator, and it is why the per-task cost column matters more in v5 than it did in v4.
No single model leads all nine evaluations. Opus 5.5 leads three — FrontierCode 2.0 at 46%, HLE-Diamond at 68% and AA-Omniscience v1.1 at 48. Astra leads Terminal-Bench-Science at 63% and the CritPt v2 hard subset at 72%. Sonnet 5.5 tops AA-Briefcase at 1,812 Elo, Kimi K3.1 Pro tops Terminal-Bench 4.0 at 67%, Gemini 4 Argon tops AutomationBench at 78%, and Muse Wonder Full tops GDP.pdf at 33% all-pass. Six different models lead six different tests, which is the most useful thing on the page: a composite hides exactly the specialisation a team is buying.
Alongside v5, Artificial Analysis launched a Progress Index — one continuous capability scale covering models released since November 2022, built with lessons from Epoch AI's Capabilities Index. It exists precisely because composite versions are not comparable across rewrites, and it is the right place to look when someone quotes a number from the summer. The September figure we reported for Astra, 61, came on an older version of the index — GPT-6 Astra scores 61 on Artificial Analysis, five behind Anthropic — and under v5 the same model measures 57 while the leader measures 58 rather than 66. Cross-version comparisons are now an error, not a shortcut.
The counter-argument writes itself. Reordering the top five in one release, in a direction that happens to favour OpenAI on coding and science, invites the accusation that the benchmarks were chosen to fit a conclusion. Practitioners made that case within hours: one widely shared reply argued Astra had never looked close to Opus 5.5 in real use and that the new index flatters it, while others pointed out that Terra now outranks Gemini 4 Argon and Fable 5.1. Nathan Lambert's read is the more durable one — hill-climbing is easy, knowing what to hill-climb is hard, and independent evaluation is under-appreciated for how hard it is to do well, as he argues in the benchmarker's dilemma. A private suite run by one organisation is more honest about contamination and less checkable by outsiders. Both things are true at once, and v5 does not resolve the tension; it relocates it.
What to watch: whether labs start quoting v5 numbers in launch posts or quietly keep citing launch-day figures from their own harnesses, and whether Kimi's promised weights, due by November 4, land close enough to the 46 that the open-weights story holds.
If a private benchmark reorders the leaderboard in one release, should readers trust the new ranking or the old one — and what would make you switch? Write to us via our contact page.
