Gap in Arena points — a public preference ranking of chat models — between the top US and top Chinese Arena rating on the LMArena text board (overall, raw ratings without style control).
US lead on this boardMeasured · LMArena dataset
US lead22 Oct 2026 (LMArena board)34.6
Arena pointsAs of 22 Oct 2026 (LMArena board)US lead narrowed 6.6 Arena points since 20 Oct
United StatesGemini 4 Argon (High)1,525.7
Arena ratingAs of 22 Oct 2026 (LMArena board)US rating ▼ 6.6 Arena points since 20 Oct
ChinaMiMo-V2.6-Pro1,491.1
Arena ratingAs of 22 Oct 2026 (LMArena board)China rating unchanged since 20 Oct
Hover or tap a point for the date and lead.
Arena rating lead, top US vs top Chinese model. Higher = US ahead.Raw LMArena ratings without style control; with style control, which Stanford uses, the US stayed about 3–7 points ahead in early 2025.Starts 9 Jan 2024, the first LMArena board with rating ranges and vote counts; earlier boards used an older rating method.Small moves may sit within the confidence interval. Updated after public releases, not live. Source: LMArena.
What this means
On this yardstick the US leads by 35 points, an implied ~55% win rate in head-to-head votes, while chips, memory, and power can still sit far apart.
Earlier figure · 2026 AI Index, March 2026. On LMArena (Stanford used its March 2026 export with style control) the US led by 39 points (1,503 vs 1,464), which Stanford reports as a 2.7% gap (39 points ÷ China's 1,464). Feb 2025: 5 points (0.4%). Different yardstick (2025 Index): end-2023 gaps of 17.5 / 13.5 / 24.3 / 31.6 pts on MMLU, MMMU, MATH, and HumanEval — not this series.
Method. LMArena dataset as republished, current vintage. Text leaderboard, category=overall, raw ratings without style control; best rating per origin; no vote floor. Gemini 4 Argon (High) 1525.7 [1519.2–1532.3] (10,083 votes) vs MiMo-V2.6-Pro 1491.1 [1481.8–1500.4] (4,067 votes). Fetched 28 Oct 2026. Changes use the rounded ratings shown, so they can differ by 0.1 from the archive's unrounded gap. Moves smaller than the bracketed rating ranges above are within LMArena's confidence interval.
Caveat. Arena standing may partly reflect platform adaptation; this is not a live leaderboard.
Open-market token share (OpenRouter top-models day sample)
Share of tokens — text models read and write — on OpenRouter, a public router. This is routed demand: usage that passed through one marketplace on one day.
China leads this sampleMeasuredOne public router · day sample
United States28.1%
of sampleAs of 27 Oct 2026US share ▼ 0.7 points since 26 Oct
China66.5%
of sampleAs of 27 Oct 2026China share ▼ 0.1 points since 26 Oct
Unmapped / stealthnever counted as China0.7%
of sampleAs of 27 Oct 2026Unmapped unchanged since 26 Oct
China ÷ USmapped China ÷ mapped US
~2.36×
As of 27 Oct 2026China ÷ US ▲ 0.05× since 26 Oct
US 28.1%·China 66.5%·Unmapped 0.7%·Rest 4.7%
OpenRouter top-models day sample — not a census
Unmapped is shown on purpose — a US–China-only bar would hide it.
What this means
China-origin models account for more of the mapped tokens than US-origin models; 0.7% of the sample is unmapped and is not counted as China.
Method. Primary cite: OpenRouter rankings (CC BY 4.0). Rounded from the day-sample extract: US 28.120% → 28.1%, China 66.454% → 66.5%, unmapped 0.702% → 0.7%, China ÷ US 2.3632 → about 2.36×. RoW 4.723% (EU 0.311% folded in) is shown as 4.7% so the bars sum to 100.0.
Caveat. Day sample of OpenRouter’s top models as of 27 Oct 2026. Excludes first-party ChatGPT/Gemini apps and China domestic routers. Unmapped or stealth models stay unmapped.
Sourced scoreboard only. 9 lanes on the board (6 Primary, 3 Estimates) plus companions on HBM wafers and the current ECI point gap — 11 tiles. Dual HBM tiles are one lane, two units (revenue and wafers). The ECI months tile and the ECI point-gap companion are one index, two units. Social posts, forums, blog rankings, and Wikipedia-only cites are not used.
Not the Primary board
Estimates
Modeled figures and packages that are hard to compare. Each tile still has the last update, one sentence of meaning, and one caveat. Not the checkable Primary board.
Current · as-of Q4 2025no update since Q4 2025
Who owns the big AI chips
Modeled share of AI computing power from ownership of leading AI chips at end-2025 — who owns them, not who uses them.
US leadsModeled — Epoch
Five US hyperscalersAmazon, Google, Meta, Microsoft, Oracle71%
of computing powerAs of Q4 2025
Chinajust over 5%~5%
of computing powerAs of Q4 2025
Five US hyperscalers 71%·China ~5%·Rest 24%
Share of AI computing power, not chip count.
What this means
A wide ownership gap can sit next to a narrow quality gap.
Caveat. The US figure is five hyperscalers — Amazon, Google, Meta, Microsoft, and Oracle — not the whole country. Excludes smuggling and offshore cloud.
Current · as-of 2 Jan 2026no update since 2 Jan 2026
Frontier capability gap in months (Epoch Capabilities Index)
Published Epoch finding of how many months China-origin models lagged the US frontier on the Epoch Capabilities Index (ECI), a modeled composite of model capability, since 2023.
Mean China lag since 2023Modeled — Epoch ECI monthsCurrent gap in months not published
Mean China lagsince 20237
ECI monthsAs of 2 Jan 2026
Minimumsince 20234
ECI monthsAs of 2 Jan 2026
Maximumsince 202314
ECI monthsAs of 2 Jan 2026
Current gapsince-2023 range only
not published — not invented
As of 2 Jan 2026
What this means
A modeled time-to-frontier average since 2023 — a different yardstick from the Arena preference ranking, and not a current gap recomputed here.
Method. Mean, minimum, and maximum are the published since-2023 range. This tile does not invent a current gap in months. The current ECI point gap is a second reading, in points.
Caveat. Published insight only (2 Jan 2026). China leaders are often open-weight, so the gap overlaps open versus closed models.
Refreshed annually after public releases. Epoch AI, 2026
Current · as-of 2024–2025no update since 2024–2025
Government industrial packages
Multi-year industrial packages, not yearly AI budgets.
Low–medium comparabilityOfficial estimateNOT COMPARABLE — not an annual AI budget race
United StatesCHIPS (39B mfg + 11B R&D)$50.0B
CHIPSAs of 2024–2025
ChinaNot annualized
Big Fund III CNY 344B (~USD 47.5B)
AI Industry Fund CNY 60.06B (~USD 8.2B)
As of 2024–2025
What this means
CHIPS and China’s funds buy different things on different clocks; they do not add to a spending race.
Caveat. Do not annualize. The US figure is CHIPS. China’s figures are Big Fund III plus the AI Industry Fund, in yuan.
Current · as-of 15 Oct 2026US–China gap unchanged since 12 Oct
Current ECI point gap
Best US minus best China ECI among models released on or before the latest scored release — ECI points, not months.
Modeled · Epoch ECI points
Current ECI point gap: US +11.0 ECI points (latest model release 15 Oct 2026; Claude Opus 5.5 vs Kimi K3)
What this means
A current point gap on Epoch's index — not the published 7-month average, and not the Arena preference ranking.
Method. The 12 Oct prior is the same Epoch file at an earlier cut-off date.
Caveat. Epoch re-fits ECI, so past points can be restated. eci_scores.csv has no Last-Modified header; the file sha256 is the vintage. This is not a gap in months.