China & the West.
The frontier AI gap.
Kimi is close to the leading models on a broad benchmark composite. The remaining gap varies sharply by task—from competitive agent performance to a substantial deficit on scientific workflows.
A snapshot of publicly reported model capabilities as of 25 October 2026. National research capacity, training compute, and unreleased systems are outside its scope.
Kimi is close to the leading composite score
Model settings, evaluation cost and scope
Selected leading Western models and Chinese comparators. Each row uses the named configuration. Index points are not percentages of intelligence. Muse Wonder Full has no public API cost listed. Cost is AA’s weighted benchmark cost per task, not training cost.
| Configuration | Lab | Index | Cost / task |
|---|
AA’s index is primarily English and text based. SciCode, 10% of the index, is currently flagged under review. Methodology.
The gap depends strongly on the task
Exact scores, comparison models and sources
| Benchmark | Kimi | Comparator | Score |
|---|
Each row compares scores under the same named evaluation, expressed as percentage points. Rows measure different abilities and are not averaged. Small differences are point estimates, not evidence of statistical significance. Independent runs use a common harness per benchmark, but reasoning compute is not equalized. Moonshot’s compilation mixes its runs and other published results with differing harnesses and budgets.
Even “knowledge work” has two different gaps
GDPval-AA v2.1
AA-Briefcase v1.1
Scores and uncertainty bounds
| Benchmark / model | Elo | Lower | Upper |
|---|
A narrow time-gap estimate: 31 days
How to interpret the timeline
All models use today’s AA v4.3.2 scores, placed at their release dates. This is a retrospective reconstruction, not an archive of what was measured at release. Incomplete or estimated composite scores are excluded. The 31-day interval measures one composite threshold; it is neither a general research lag nor a catch-up forecast. Lab location defines the regional groups, not the origin of individual researchers or hardware.
| Release date | Frontier model | Index |
|---|
Read the measurements at their actual scale.
Evidence and comparability
The default views use independent Artificial Analysis evaluations. The alternate benchmark view uses Moonshot’s launch compilation, which mixes its own runs with published peer results. Tools, harnesses, reasoning settings, and compute budgets can change the ordering.
Coverage and uncertainty
The AA v4.3.2 composite is primarily English and text based. SciCode, 10% of its weight, is flagged under review. Some evaluations use private tasks. Small score differences are point estimates; reported uncertainty intervals are shown where available.
Geography and time
Regional groups follow lab location. “West” covers the US, Canada, France, Spain, and Switzerland in the source data. The timeline excludes incomplete or estimated composite scores and applies the current suite retrospectively. It is not an archive of release-day rankings.