Title: AI Model Weekly Rankings: Google's Gemini 4 Carbon Debuts Strong at No. 1, Chinese Models Hold Firm in Video Top-Five Tier
Body:
AI Model Weekly Rankings: Google's Gemini 4 Carbon Debuts Strong at No. 1, Chinese Models Hold Firm in Video Top-Five Tier
IT Home, November 16: The Arena (arena.ai) platform has released its blind-test voting rankings for large models for week 46 of 2026 (November 9 to November 15, 2026).
The biggest highlight this week comes from Google: the new model Gemini 4 Carbon (High) debuted at No. 1 on the overall leaderboard with 1526 points, followed closely by its sibling Gemini 4 Argon (High) with 1522 points. The two differ by only 4 points and are considered tied for first within the margin of error.
On the Chinese-model front, on the frontend development leaderboard, Moonshot's kimi k3.1 pro (Max) entered the rankings for the first time at No. 12, becoming the highest-ranked Chinese model on that list; on the AI video leaderboard, ByteDance's dreamina seedance 2.0 / 2.5 took No. 6 and No. 7 respectively, while Alibaba's wan 3.0 placed No. 8, with the Chinese contingent firmly holding a place in the top tier of the video generation track.
Overall, this week's rankings present a landscape of "a new Google king on the throne, Anthropic holding firm in agents and frontend, and Chinese models continuing to penetrate multimodal tracks."
IT Home note: This leaderboard is based on anonymous blind-test voting on the Arena (arena.ai) platform, with rankings calculated via ELO (the agent leaderboard uses a percentage signal). It reflects users' subjective preferences rather than absolute capability tests; the different sub-leaderboards are independent of one another and should not be compared across boards, score gaps within the margin of error are treated as ties, and new models must accumulate a certain number of votes before they formally enter the rankings.
Overall Leaderboard
The overall leaderboard saw significant changes this week. Google's newly launched Gemini 4 Carbon (High) landed at No. 1 with 1526 points, followed by its older sibling Gemini 4 Argon (High) at No. 2 with 1522 points. The gap is only 4 points, treated as a tie at the top within the confidence interval.
Meta's Muse Wonder (Max) placed No. 3 with 1511 points, and Anthropic's Claude Opus 5.5 (High) placed No. 4 with 1507 points.
Notably, Anthropic holds 5 of the Top 8 spots (Claude Opus 5.5, Claude Fable 5, Claude Opus 4.6, Claude Fable 5.1, Claude Opus 4.7), a clear family-style occupation, though the score gaps are extremely small, all within 6 points. Claude Opus 4.8 (High) placed No. 27 with 1482 points, the model with the most obvious decline in the Anthropic camp, overtaken by several models in the same tier.
Chinese models sit in the midrange of the overall leaderboard, still about 40 points behind the leaders:
- Moonshot Kimi K3.1 Pro (Max) ranks highest, at No. 24 with an ELO of 1485;
- Moonshot Kimi K3 (Max) placed No. 20 with an ELO of 1487;
- Alibaba Qwen 3.8 Max placed No. 25 with an ELO of 1484;
- Zhipu GLM 5.3 (Max) placed No. 28 with an ELO of 1482.
| Rank | Model | Vendor | Score (±CI) | Votes | Price ($/M) Input / Output | Context |
|---|---|---|---|---|---|---|
| 1 | Gemini 4 Carbon (High) | 1526 ±10 | 4273 | $$2 / $10 | N/A | |
| 2 | Gemini 4 Argon (High) | 1522 ±5 | 20779 | $$2 / $10 | 1M | |
| 3 | Muse Wonder (Max) | Meta | 1511 ±6 | 12696 | $$4 / $20 | 1M |
| 4 | Claude Opus 5.5 (High) | Anthropic | 1507 ±6 | 15793 | $$4 / $20 | 1M |
| 5 | Claude Fable 5 (High) | Anthropic | 1504 ±4 | 45469 | $$10 / $50 | 1M |
| 6 | Claude Opus 4.6 (High) | Anthropic | 1503 ±3 | 84375 | $$5 / $25 | 1M |
| 7 | Claude Fable 5.1 (Max) | Anthropic | 1502 ±5 | 20511 | $$10 / $50 | 1M |
| 8 | Claude Opus 4.7 (High) | Anthropic | 1501 ±4 | 64260 | $$5 / $25 | 1M |
| 9 | Gemini 3.8 Flash (High) | 1497 ±4 | 62675 | $$0.75 / $3.75 | 1M | |
| 10 | Muse Spark 1.3 (Max) | Meta | 1497 ±5 | 26806 | $$1.25 / $4.25 | 1M |
| — Below are Chinese models outside the Top 10 — | ||||||
| 20 | Kimi K3 (Max) | Moonshot | 1487 ±4 | 41402 | N/A | N/A |
| 24 | Kimi K3.1 Pro (Max) | Moonshot | 1485 ±10 | 3995 | $$2 / $10 | 1M |
| 25 | Qwen 3.8 Max | Alibaba | 1484 ±5 | 36936 | $$2 / $6 | 1M |
| 28 | GLM 5.3 (Max) | Z.ai | 1482 ±5 | 31604 | $$1.40 / $4.40 | 1M |
Frontend Development Leaderboard
The frontend development leaderboard this week is strongly dominated by Anthropic. Claude Fable 5.2 (xHigh) sits at No. 1 with 1838 points, 23 points ahead of No. 2 Claude Opus 5.5 (Max) (1815 points), a clear advantage.
Anthropic holds 3 of the Top 5 spots (Claude Fable 5.2, Claude Opus 5.5, Claude Sonnet 5.5), while Google's Gemini 4 Carbon (High) placed No. 4 with 1768 points and OpenAI GPT 6 Astra (Max) placed No. 5 with 1764 points.
Chinese models performed impressively on this leaderboard: Moonshot's kimi k3.1 pro (Max) placed No. 12 with 1687 points, the highest-ranked Chinese model, breaking directly into the top 15; Alibaba, DeepSeek, Baidu, Xiaomi, Tencent, and Zhipu all have models in the top 30, with the Chinese contingent taking 8 spots in the top 30 — the most seats for Chinese models on any of the sub-leaderboards.
AI Large Model Weekly Leaderboard: Google Gemini 4 Carbon Makes Strong Debut at No. 1; Domestic Models Hold Firm in Video Leaderboard Top-Five Camp
| Rank | Model | Provider | Score (±CI) | Votes | Price ($/M) Input / Output | Context |
|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 (xHigh) | Anthropic | 1838 ±21 | 1215 | $$8 / $40 | 1M |
| 2 | Claude Opus 5.5 (Max) | Anthropic | 1815 ±12 | 4208 | $$4 / $20 | 1M |
| 3 | Claude Sonnet 5.5 (xHigh) | Anthropic | 1772 ±10 | 5895 | $$2 / $10 | 1M |
| 4 | Gemini 4 Carbon (High) | 1768 ±15 | 2120 | $$2 / $10 | N/A | |
| 5 | GPT 6 Astra (Max) | OpenAI | 1764 ±9 | 8470 | $$10 / $50 | 1.1M |
| 6 | GPT 6.1 Sol (Max) | OpenAI | 1743 ±10 | 5590 | $$2 / $10 | 1.1M |
| 7 | Claude Fable 5.1 (Max) | Anthropic | 1742 ±8 | 9800 | $$10 / $50 | 1M |
| 8 | Muse Wonder Full (Max) | Meta | 1723 ±14 | 2068 | N/A | N/A |
| 9 | Claude Sonnet 5.5 (High) | Anthropic | 1715 ±8 | 7940 | $$2 / $10 | 1M |
| 10 | Gemini 4 Hydrogen (High) | 1704 ±14 | 1997 | $$1.50 / $7.50 | N/A | |
| — Below are domestic models outside the Top 10 — | ||||||
| 12 | Kimi K3.1 Pro (Max) | Moonshot | 1687 ±10 | 4977 | $$2 / $10 | 1M |
| 15 | Qwen 3.8 (Max 0902) | Alibaba | 1676 ±6 | 22099 | $$2 / $6 | N/A |
| 17 | Qwen 3.8 Max | Alibaba | 1670 ±12 | 3370 | $$2 / $6 | 1M |
| 18 | DeepSeek V4.1 Flash (Max) (1017) | DeepSeek | 1664 ±7 | 14108 | $$0.15 / $0.60 | 1M |
| 18 | Ernie 5.5 Flash Preview | Baidu | 1668 ±16 | $$0.28 / $1.13 | N/A | |
| 19 | Kimi K3 (Max) | Moonshot | 1658 ±6 | 22448 | $$3 / $15 | 1M |
| 25 | Mimo V2.6 Flash | Xiaomi | 1638 ±8 | 8944 | $$0.14 / $0.28 | 1.1M |
| 26 | Hy4 (Preview) | Tencent | 1632 ±10 | 4601 | $$0.83 / $2.50 | 1M |
| 28 | Mimo V2.6 Pro | Xiaomi | 1625 ±8 | 8846 | $$0.43 / $0.87 | 1.1M |
| 29 | Qwen 3.8 Flash Next | Alibaba | 1624 ±7 | 11380 | $$0.16 / $0.47 | 1M |
Agent Leaderboard
The Agent Leaderboard this week is completely dominated by Anthropic. Claude Fable 5.2 (Max) ranks No. 1 with a 30.48% net improvement, more than 11 percentage points ahead of No. 2 Claude Opus 5.5 (Max) (19.51%), a significant advantage.
Google Gemini 4 Carbon (High) ranks No. 3 with 16.9%, and its tool hallucination rate is only 0.55%, making it one of the "cleanest" in tool calling among leading models.
Anthropic occupies 5 of the top 8, OpenAI has one each at No. 7, 9, and 10, and Google has one each at No. 3, 5, and 11.
Among domestic models, Tencent Hy4 ranks No. 18 with 5.75%, and Moonshot Kimi K3.1 Pro (Max) ranks No. 20 with 4.83%, at the tail end of the second tier.
| Rank | Model | Provider | Net Improvement (±CI) | Task Confirmation Success Rate | Positive / Negative Review Ratio | Controllability |
|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 (Max) | Anthropic | 30.48% ±3.07% | 0% ±0% | 74.41% ±12.7% | 32.91% ±1.41% |
| 2 | Claude Opus 5.5 (Max) | Anthropic | 19.51% ±2.04% | 0% ±0% | 40.72% ±7.59% | 18.41% ±1.94% |
| 3 | Gemini 4 Carbon (High) | 16.9% ±0.68% | 0% ±0% | 40.91% ±2.48% | 19.06% ±0.61% | |
| 4 | Claude Sonnet 5.5 (Max) | Anthropic | 16.07% ±2.81% | 0% ±0% | 31.11% ±11% | 14.53% ±2.73% |
| 5 | Gemini 4 Hydrogen (High) | 14.8% ±2.23% | 0% ±0% | 32.12% ±8.39% | 20.53% ±2.04% | |
| 6 | Claude Fable 5.1 (Max) | Anthropic | 14.62% ±2.06% | 0% ±0% | 31.14% ±7.71% | 14.49% ±1.85% |
| 7 | GPT-6 Astra (Max) | OpenAI | 14.56% ±2.03% | 0% ±0% | 31.86
AI Large Model Weekly Ranking: Google Gemini 4 Carbon Makes a Strong Debut at No. 1, Domestic Models Remain Firmly in the Video Ranking's Top-Five Camp
| Rank | Model | Vendor | Score (±CI) | Votes | Price ($/M) Input / Output | Context |
|---|---|---|---|---|---|---|
| 1 | GPT Image 2.5 Sunburst | OpenAI | 1429 ±5 | 63818 | ||
| 2 | GPT Image 2.6 Sunburst (Medium) | OpenAI | 1406 ±13 | 2697 | ||
| 3 | GPT Image 2.5 Flare | OpenAI | 1398 ±5 | 63120 | ||
| 4 | GPT Image 2.6 Flare (Medium) | OpenAI | 1384 ±13 | 2532 | ||
| 5 | GPT Image 2 (medium) | OpenAI | 1382 ±4 | 108559 | ||
| 6 | Mai Image 2.6 | Microsoft AI | 1330 ±5 | 41413 | ||
| 7 | Gemini Nano Banana 2.1 | 1327 ±9 | 4961 | |||
| 8 | Grok Imagine Image 2.0 (canvas) | SpaceXAI | 1318 ±7 | 9919 | ||
| 9 | Reve 2.1 | Reve | 1301 ±8 | 7893 | ||
| 10 | Grok Imagine Image 2.0 (20260801) | SpaceXAI | 1287 ±5 | 36215 | ||
| — Domestic models outside the Top 10 below — | ||||||
| 14 | Qwen Image 3.0 Pro | Alibaba | 1254 ±7 | 10899 | ||
| 17 | Seedream 5.0 Pro | Bytedance | 1248 ±4 | 110580 | ||
| 21 | Qwen Image 2.1 | Alibaba | 1227 ±6 | 17822 | ||
| 23 | MiniMax H3 Image | MiniMax | 1197 ±11 | 3277 | ||
| 24 | Qwen Image 2.0 Pro 2026.06.22 | Alibaba | 1190 ±6 | 12657 |
AI Video Ranking
The AI Video Ranking this week shows a pattern of "Google leading, domestic models closely chasing."
Google's Gemini Omni 1.2 Flash debuted at No. 1 with 1540 points, while its predecessor Gemini Omni 1.1 Flash ranked No. 2 with 1513 points. The 27-point gap between them remains within the 30-point margin of error, so they can be regarded as tied at the top.
Domestic models performed exceptionally well: ByteDance's Dreamina Seedance 2.0 720p ranked No. 6 with 1483 points, Dreamina Seedance 2.5 720p ranked No. 7 with 1479 points, and Alibaba's Wan 3.0 ranked No. 8 with 1476 points. The domestic camp holds 3 of the top 10 spots in the video ranking, making it the highest-ranked and most concentrated among all sub-rankings.
Notably, OpenAI's Sora 2 Pro ranked only No. 13 with 1367 points, overtaken by multiple domestic video models.
| Rank | Model | Vendor | Score (±CI) | Votes | Price ($/M) Input / Output | Context |
|---|---|---|---|---|---|---|
| 1 | Gemini Omni 1.2 Flash | 1540 ±16 | 1714 | |||
| 2 | Gemini Omni 1.1 Flash | 1513 ±15 | 1792 | |||
| 3 | Gemini Omni Flash | 1509 ±8 | 46707 | |||
| 4 | Flux 3 Video | Black Forest Labs | 1491 ±17 | 1314 | ||
| 5 | Grok Imagine Video 1.5 (Agent) | SpaceXAI | 1488 ±19 | 1249 | ||
| 6 | Dreamina Seedance 2.0 720p | Bytedance | 1483 ±8 | 72137 | ||
| 7 | Dreamina Seedance 2.5 720p | Bytedance | 1479 ±8 | 25421 | ||
| 8 | Wan 3.0 | Alibaba | 1476 ±12 | 3060 | ||
| 9 | SkyReels V4 | Skywork | 1459 ±18 | 1060 | ||
| 10 | Muse Video | Meta | 1457 ±15 | 2225 | ||
| — Domestic models outside the Top 10 below — | ||||||
| 11 | Minimax H3 | MiniMax | 1456 ±8 | 29257 | ||
| 12 | Happyhorse 1.0 | Alibaba-ATH | 1429 ±12 | 22896 | ||
| 22 | Wan 2.7 | Alibaba | 1337 ±7 | 48517 | ||
| 23 | Wan 2.6 | Alibaba | 1324 ±7 | 75753 | ||
| 24 | Seedance V1.5 Pro | Bytedance | 1256 ±7 | 77992 | ||
| 27 | Wan 2.5 I2i (Preview) | Alibaba | 1244 ±9 | 43113 | ||
| 28 | Pixverse V5.6 | Proprietary | 1238 ±10 | 33302 | ||
| 30 | Kling 2.5 Turbo 1080p | KlingAI | 1219 ±17 | 2110 |
Summary
Looking across the six major rankings this week, the competitive landscape for AI large models shows three clear main threads:
- First, with the release of Gemini 4 Carbon, Google has pushed simultaneously in the overall ranking