China’s frontier model is close.
The gap is not uniform.
Kimi K3.1 Pro sits three points behind the Western leader on an independent composite. It is competitive in agentic coding and long context, while Western models retain a broader edge in knowledge reliability, science and reasoning.
A small aggregate deficit—not an across-the-board lag. Kimi belongs to the current frontier cohort, but “best-of-West” still wins most individual capability comparisons.
Independent runs put Kimi ahead on Terminal-Bench 4.0 (+3 points) and AA-LCR long-context reasoning (+1). It also outranks GPT-6 Astra, Gemini 4 Argon and several other Western models on the aggregate index.
Kimi trails the best Western result on the other eight index components. Gaps include AA-Omniscience (−16 index points), Humanity’s Last Exam (−9 points), CritPt (−8) and SciCode (−6).
1 · The aggregate frontier
The independent composite puts Kimi three index points behind Claude Opus 5.5. This is a snapshot of published model performance, not a measurement of national research capacity. The previous Kimi K3 is included only to show the size of Moonshot’s latest jump.
2 · Independent component results
This comparison uses the same third-party evaluator for every model. “Best Western” is the maximum among Opus 5.5, Sonnet 5.5, Muse Wonder Full, Fable 5.1, GPT-6 Astra, Gemini 4 Argon, GPT-6.1 Sol and Muse Wonder, selected separately for each component. It is a demanding composite, not one deployable model.
3 · Moonshot’s published benchmark panel
Moonshot’s own selection looks closer: on the 12 public percentage-scale tests, Kimi has 6 leads and 6 deficits, with a median signed gap of −0.15 percentage points. The main exception is Terminal-Bench Science, where Kimi trails GPT-6 Astra by 30 points.
How to read—and not read—this
- Scope: “China” is represented by one Moonshot model, not the full Chinese ecosystem. “West” is represented by the named US models; it is not a complete sample of Western labs.
- No time-lag conversion: a three-point index gap does not identify a number of months or years behind. That would require a separate historical analysis.
- No percentage conversion: 55 vs 58 is not “95% as intelligent.” The index is a weighted composite, not a ratio scale of general intelligence.
- Different evidence: Chart 2 uses independent Artificial Analysis runs. Chart 3 uses Moonshot’s vendor-curated compilation, whose footnotes mix evaluation harnesses and score provenance. Do not compare their raw values directly.
- Selection matters: Moonshot’s internal Kimi Code Bench is excluded from Chart 3. GDPval-AA Elo is reported separately rather than mixed with percentage-point gaps. Some Artificial Analysis evaluations are proprietary even though their results are public.
- Small differences are not necessarily meaningful: these charts do not estimate sampling error or statistical significance. Treat narrow leads as approximate parity.
Method and source pages
The overall index is Artificial Analysis Intelligence Index v4.3.2, which combines AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1. See the index methodology.
Primary sources
Independent model comparisons used for best-of-West