
IT之家 10 月 19 日消息,Arena(arena.ai)平台发布 2026 年第 42 周( 2026-10-12 ~ 2026-10-18 )AI 大模型盲测排名,本期头部格局总体稳定,国产模型在多个细分榜单保持稳定中上游位置,其中 kimi-k3-max 首次稳定在代码榜前十,mimo-v2.6-pro 首次进入前端开发榜前 20,总体表现可圈可点。本期共有三款新模型正式入榜,分别是 gemini-3.5-flash-medium、 mimo-v2.6-pro 以及 flux-2-flex。
IT之家注:本榜单基于 Arena(arena.ai)平台匿名盲测投票,借助 ELO(智能体榜为百分比信号)计算排名,反映使用者主观偏好而非绝对能力测试;不同分榜相互独立,不宜跨榜比较,分差在误差范围内均视为并列,新模型也需积累一定投票量后才会正式入榜。
综合榜
本期综合榜头部排名保持稳定,gemini-4-argon-high 以 1525 分蝉联榜首,claude-opus-5.5-high 以 1507 分位列第二,claude-opus-4-6-high 和 claude-fable-5-high 同以 1504 分并列第三,分差均在 30 分以内,在误差范围内视为接近竞争。头部前 10 名分数均在 1494 分以上,竞争依旧胶着。本周新入榜模型为 gemini-3.5-flash-medium,直接冲进前 30,位列第 30,ELO 得分 1476 分,表现不俗。
国产模型总体处于中上游位置,梯队相对稳定:
| 排名 | 较上周 | 模型 | 厂商 | 分数 (±CI) | 票数 | 售价 ($/M) 输入 / 输出 | 上下文 |
|---|---|---|---|---|---|---|---|
| 1 | - | gemini-4-argon-high | 1525 ±9 | 4892 | $2 / $10 | 1M | |
| 2 | - | claude-opus-5.5-high | Anthropic | 1507 ±8 | 6272 | $4 / $20 | 1M |
| 3 | - | claude-opus-4-6-high | Anthropic | 1504 ±3 | 79415 | $5 / $25 | 1M |
| 4 | - | claude-fable-5-high | Anthropic | 1504 ±4 | 40026 | $10 / $50 | 1M |
| 5 | - | claude-opus-4-7-high | Anthropic | 1501 ±4 | 65192 | $5 / $25 | 1M |
| 6 | - | claude-fable-5.1-max | Anthropic | 1501 ±6 | 13063 | $10 / $50 | 1M |
| 7 | - | claude-opus-4-6 | Anthropic | 1498 ±3 | 84095 | $5 / $25 | 1M |
| 8 | - | gemini-3.8-flash-high | 1497 ±5 | 31437 | $0.75 / $3.75 | 1.05M | |
| 9 | - | muse-spark-1.3-max | Meta | 1494 ±6 | 14694 | $1.25 / $4.25 | 1M |
| 10 | - | claude-opus-4-7 | Anthropic | 1494 ±4 | 66315 | $5 / $25 | 1M |
| — 以下为 Top 10 外的国产模型 — | |||||||
| 16 | - | kimi-k3-max | Moonshot | 1488 ±5 | 30351 | N/A | N/A |
| 22 | - | qwen3.8-max | Alibaba | 1483 ±5 | 25378 | $2 / $6 | 1M |
| 26 | - | mimo-v2.6-pro | Xiaomi | 1480 ±9 | 4069 | $0.44 / $0.87 | 1.05M |
| 27 | - | glm-5.3-max | Z.ai | 1478 ±5 | 20049 | $1.4 / $4.4 | 1.05M |
代码榜
代码榜头部格局依旧稳定,gemini-4-argon-high 以 1562 分继续领跑,claude-fable-5-high 以 1551 分位列第二,claude-opus-4-7-high、 claude-opus-4-6-high、 claude-opus-5.5-high 同以 1550 分并列第三,前三名中除 gemini-4-argon-high 外均为 Anthropic 模型,头部依旧被海外厂商占据。本周核心看点是国产 kimi-k3-max 稳定保持在第十名,ELO 1540 分,成为国产模型在代码榜的最高排名,小米 mimo-v2.6-pro 紧随其后位列第十一,ELO 1539 分,表现亮眼。
国产模型已有多款进入代码榜前 30,具体排名如下:
| 排名 | 较上周 | 模型 | 厂商 | 分数 (±CI) | 票数 | 售价 ($/M) 输入 / 输出 | 上下文 |
|---|---|---|---|---|---|---|---|
| 1 | - | gemini-4-argon-high | 1562 ±17 | 1233 | $2 / $10 | 1M | |
| 2 | - | claude-fable-5-high | Anthropic | 1551 ±7 | 10404 | $10 / $50 | 1M |
| 3 | - | claude-opus-4-7-high | Anthropic | 1550 ±6 | 18591 | $5 / $25 | 1M |
| 4 | - | claude-opus-4-6-high | Anthropic | 1550 ±6 | 20617 | $5 / $25 | 1M |
| 5 | - | claude-opus-5.5-high | Anthropic | 1550 ±16 | 1444 | $4 / $20 | 1M |
| 6 | - | claude-opus-4-6 | Anthropic | 1548 ±5 | 23351 | $5 / $25 | 1M |
| 7 | - | claude-opus-4-7 | Anthropic | 1546 ±6 | 18816 | $5 / $25 | 1M |
| 8 | - | gpt-6.1-sol-max | OpenAI | 1545 ±20 | 911 | $2 / $10 | 1.05M |
| 9 | - | gpt-6-astra-max | OpenAI | 1542 ±12 | 2424 | N/A | N/A |
| 10 | - | kimi-k3-max | Moonshot | 1540 ±7 | 7775 | N/A | N/A |
| — 以下为 Top 10 外的国产模型 — | |||||||
| 11 | - | mimo-v2.6-pro | Xiaomi | 1539 ±18 | 1107 | $0.44 / $0.87 | 1.05M |
| 26 | - | deepseek-v4.1-flash-max | DeepSeek | 1528 ±11 | 3026 | $0.3 / $1.2 | 1M |
| 27 | - | glm-5.3-flash | Z.ai | 1525 ±8 | 7267 | $0.06 / $0.2 | 1.05M |
| 28 | - | qwen3.7-max-preview | Alibaba | 1524 ±18 | 1165 | $1.48 / $4.43 | 1M |
| 29 | - | qwen3.8-max | Alibaba | 1524 ±8 | 7077 | $2 / $6 | 1M |
前端开发榜
前端开发榜头部格局稳定,claude-opus-5.5-max 以 1813 分稳居榜首,gpt-6-astra-max 以 1786 分位列第二,claude-sonnet-5.5-xhigh 以 1774 分位列第三。本周新入榜模型涵盖 mimo-v2.6-pro 和 claude-haiku-5.5-high,其中小米 mimo-v2.6-pro 首次入榜即位列第 19,ELO 1629 分,成功杀入前 20,延续了其在代码领域的良好表现。
国产模型在前端开发榜占据多个中上游位置,具体如下:
| 排名 | 较上周 | 模型 | 厂商 | 分数 (±CI) | 票数 | 售价 ($/M) 输入 / 输出 | 上下文 |
|---|---|---|---|---|---|---|---|
| 1 | - | claude-opus-5.5-max | Anthropic | 1813 ±15 | 2294 | $4 / $20 | 1M |
| 2 | - | gpt-6-astra-max | OpenAI | 1786 ±11 | 5189 | $10 / $50 | 1.05M |
| 3 | - | claude-sonnet-5.5-xhigh | Anthropic | 1774 ±15 | 2323 | $2 / $10 | 1M |
| 4 | - | gpt-6.1-sol-max | OpenAI | 1755 ±15 | 2102 | $2 / $10 | 1.05M |
| 5 | - | claude-fable-5.1-max | Anthropic | 1744 ±9 | 6663 | $10 / $50 | 1M |
| 6 | - | claude-sonnet-5.5-high | Anthropic | 1715 ±11 | 3984 | $2 / $10 | 1M |
| 7 | - | claude-opus-5-max | Anthropic | 1691 ±6 | 17320 | $5 / $25 | 1M |
| 8 | - | gpt-6-sol-max | OpenAI | 1688 ±12 | 3407 | $2 / $10 | 1.05M |
| 9 | - | gemini-4-argon-high | 1678 ±15 | 1771 | $2 / $10 | 1M | |
| 10 | - | qwen3.8-max-0902 | Alibaba | 1674 ±8 | 11182 | $2 / $6 | N/A |
| — 以下为 Top 10 外的国产模型 — | |||||||
| 11 | - | qwen3.8-max | Alibaba | 1672 ±12 | 3446 | $2 / $6 | 1M |
| 14 | - | kimi-k3-max | Moonshot | 1654 ±6 | 17412 | $3 / $15 | 1.05M |
| 16 | - | mimo-v2.6-flash | Xiaomi | 1637 ±10 | 4058 | $0.14 / $0.28 | 1.05M |
| 17 | - | qwen3.8-flash-next | Alibaba | 1633 ±8 | 6932 | $0.16 / $0.47 | 1M |
| 18 | - | hy4-preview | Tencent | 1632 ±10 | 4792 | $0.75 / $2.25 | 1.05M |
| 19 | 新上榜 | mimo-v2.6-pro | Xiaomi | 1629 ±11 | 3649 | $0.44 / $0.87 | 1.05M |
| 22 | - | glm-5.3-max | Z.ai | 1622 ±8 | 8382 | $1.4 / $4.4 | 1.05M |
| 23 | - | deepseek-v4.1-flash-max | DeepSeek | 1619 ±11 | 3542 | $0.3 / $1.2 | 1M |
| 26 | - | glm-5.3-flash | Z.ai | 1609 ±7 | 13697 | $0.06 / $0.2 | 1.05M |
| 27 | - | glm-5.2-max | Z.ai | 1603 ±6 | 15351 | $1.4 / $4.4 | 1M |
智能体榜
智能体榜使用百分比信号评分体系,本周 Claude Opus 5.5 (Max) 以 15.53% 的净提高度蝉联榜首,GPT 6 Astra (Max) 排名上升一位至第二,净提高度 13.09%,Claude Fable 5.1 (Max) 位列第三,净提高度 12.66%,头部依旧被海外模型占据。在指标维度,Claude Opus 5.5 (Max) 工具幻觉率仅 0.37%,为头部最低,表现出色。国产模型总体处于中上游梯队,已有多款进入前 30。
国产模型在智能体榜具体表现如下:
| 排名 | 较上周 | 模型 | 厂商 | 净提高度 (±CI) | 任务确认成功率 | 好评 / 差评比 | 可控性 |
|---|---|---|---|---|---|---|---|
| 1 | - | Claude Opus 5.5 (Max) | Anthropic | 15.53% ±4.63% | 17.29% ±6.02% | 36.78% ±20.55% | 8.2% ±4.06% |
| 2 | ↑ 1 | GPT 6 Astra (Max) | OpenAI | 13.09% ±2.37% | 12.13% ±3.88% | 38.54% ±9.96% | 8.85% ±2.5% |
| 3 | ↑ 1 | Claude Fable 5.1 (Max) | Anthropic | 12.66% ±2.05% | 16.19% ±2.96% | 26.22% ±7.98% | 8.84% ±3.12% |
| 4 | ↑ 1 | Claude Sonnet 5.5 (Max) | Anthropic | 11.95% ±3.31% | 15.83% ±5.06% | 23.37% ±14.09% | 5.1% ±3.33% |
| 5 | ↑ 1 | GPT 6.1 Sol (Max) | OpenAI | 11.72% ±3.05% | 14.74% ±4.18% | 30.32% ±12.99% | 8.52% ±3.07% |
| 6 | ↑ 1 | GPT 6 Sol (Max) | OpenAI | 9.92% ±2.38% | 6.97% ±4.34% | 23.55% ±9.64% | 11.92% ±2.02% |
| 7 | ↑ 1 | Gemini 4 Argon (High) | 9.26% ±2.36% | 15.64% ±3.26% | 40.6% ±9.56% | 11.64% ±2.1% | |
| 8 | ↑ 1 | Claude Fable 5 (High) | Anthropic | 8.94% ±1.27% | 5.79% ±2.6% | 19.34% ±4.8% | 9.49% ±1.39% |
| 9 | ↑ 1 | Claude Opus 5 (Max) | Anthropic | 8.12% ±1.47% | 10.42% ±3.1% | 11.86% ±5.46% | 4.22% ±2.01% |
| 10 | ↑ 2 | Claude Opus 4.8 (High) | Anthropic | 7.13% ±1.37% | 4.68% ±2.75% | 15.14% ±5.03% | 8.41% ±1.6% |
| — 以下为 Top 10 外的国产模型 — | |||||||
| 15 | ↑ 2 | Deepseek V4.1 Flash (Max) | DeepSeek | 3.8% ±0.39% | 7.44% ±0.86% | 2.44% ±1.45% | 0.87% ±0.54% |
| 16 | ↑ 2 | Kimi K3 (Max) | Moonshot | 3.78% ±0.59% | 6.46% ±1.32% | 6.44% ±2.05% | 1.14% ±0.86% |
| 17 | ↑ 2 | Hy4 preview | Tencent | 3.61% ±0.7% | 5.96% ±1.48% | 5.18% ±2.59% | 0.32% ±0.92% |
| 18 | ↑ 2 | MiMo V2.6 Pro | Xiaomi | 3.52% ±1.92% | 8.03% ±3.88% | 8.16% ±7.31% | 0.56% ±2.45% |
| 19 | ↑ 2 | GLM 5.2 (Max) | Z.ai | 3.19% ±0.73% | 4.11% ±1.58% | 8.94% ±2.61% | 1.72% ±0.96% |
| 22 | ↑ 2 | Qwen3.8 Max | Alibaba | 2.35% ±0.59% | 5.09% ±1.32% | 2.95% ±2.13% | 1.81% ±0.82% |
| 23 | ↑ 2 | GLM 5.3 (Max) | Z.ai | 2.15% ±0.72% | 4.78% ±1.62% | 5.12% ±2.48% | 0.44% ±1.08% |
| 25 | ↑ 2 | DeepSeek V4 Pro (High) (0813) | DeepSeek | 1.34% ±1.66% | 2% ±4.01% | 1.86% ±5.99% | 2.12% ±2.2% |
AI 生图榜
AI 生图榜头部格局稳定,gpt-image-2.5-sunburst 以 1425 分蝉联榜首,gpt-image-2.5-flare 以 1398 分位列第二,gpt-image-2 (medium) 以 1383 分位列第三,前三名均为 OpenAI 模型。本周新入榜模型为 flux-2-flex,首次入榜即位列第 30,ELO 得分 1157 分。国产模型在该榜表现稳定,字节跳动 seedream-5.0-pro 和阿里 qwen-image-3.0-pro 同以 1255 分并列第 12、13 名,均进入前 15,处于中上游位置。
| 排名 | 较上周 | 模型 | 厂商 | 分数 (±CI) | 票数 | 售价 ($/M) 输入 / 输出 | 上下文 |
|---|---|---|---|---|---|---|---|
| 1 | - | gpt-image-2.5-sunburst | OpenAI | 1425 ±7 | 17752 | N/A | N/A |
| 2 | - | gpt-image-2.5-flare | OpenAI | 1398 ±7 | 16889 | N/A | N/A |
| 3 | - | gpt-image-2 (medium) | OpenAI | 1383 ±4 | 93334 | $8 / $30 | N/A |
| 4 | - | grok-imagine-image-2.0 (canvas) | SpaceXAI | 1334 ±12 | 3113 | N/A | N/A |
| 5 | - | mai-image-2.6 | Microsoft AI | 1332 ±5 | 25091 | N/A | N/A |
| 6 | - | gemini-nano-banana-2.1 | 1328 ±9 | 5310 | $1.5 / $7.5 | 65.5K | |
| 7 | - | reve-2.1 | Reve | 1302 ±8 | 8249 | N/A | N/A |
| 8 | - | grok-imagine-image-2.0 (20260801) | SpaceXAI | 1294 ±6 | 14483 | N/A | N/A |
| 9 | - | muse-image | Meta | 1274 ±5 | 42262 | N/A | N/A |
| 10 | - | reve-2.0 | Reve | 1269 ±6 | 15773 | N/A | N/A |
| — 以下为 Top 10 外的国产模型 — | |||||||
| 12 | - | seedream-5.0-pro | Bytedance | 1255 ±4 | 94181 | N/A | N/A |
| 13 | - | qwen-image-3.0-pro | Alibaba | 1255 ±6 | 11440 | N/A | N/A |
| 19 | - | qwen-image-2.1 | Alibaba | 1223 ±7 | 11441 | N/A | N/A |
| 21 | - | qwen-image-2.0-pro-2026-06-22 | Alibaba | 1192 ±6 | 13096 | N/A | N/A |
AI 视频榜
AI 视频榜头部依旧由谷歌模型占据,gemini-omni-1.1-flash 以 1516 分位列第一,gemini-omni-flash 以 1513 分位列第二,flux-3-video 以 1493 分位列第三。国产模型表现突出,前五名中占据两席,字节跳动 dreamina-seedance-2.0-720p 位列第五,ELO 1479 分,阿里 wan3.0 紧随其后位列第六,ELO 1476 分,前 10 名中共有四款国产模型,总体实力不容小觑。
| 排名 | 较上周 | 模型 | 厂商 | 分数 (±CI) | 票数 | 售价 ($/M) 输入 / 输出 | 上下文 |
|---|---|---|---|---|---|---|---|
| 1 | - | gemini-omni-1.1-flash | 1516 ±15 | 1784 | $1.5 / $9 | N/A | |
| 2 | - | gemini-omni-flash | 1513 ±9 | 26576 | N/A | N/A | |
| 3 | - | flux-3-video | Black Forest Labs | 1493 ±17 | 1302 | N/A | N/A |
| 4 | - | grok-imagine-video-1.5-agent | SpaceXAI | 1492 ±18 | 1218 | N/A | N/A |
| 5 | - | dreamina-seedance-2.0-720p | Bytedance | 1479 ±8 | 56318 | N/A | N/A |
| 6 | - | wan3.0 | Alibaba | 1476 ±13 | 2598 | N/A | N/A |
| 7 | - | dreamina-seedance-2.5-720p | Bytedance | 1474 ±9 | 8001 | N/A | N/A |
| 8 | - | minimax-h3 | MiniMax | 1460 ±9 | 11417 | N/A | N/A |
| 9 | - | muse-video | Meta | 1456 ±15 | 2188 | N/A | N/A |
| 10 | - | happyhorse-1.0 | Alibaba-ATH | 1427 ±13 | 22264 | N/A | N/A |
| — 以下为 Top 10 外的国产模型 — | |||||||
| 20 | - | wan2.7-t2v | Alibaba | 1337 ±7 | 27626 | N/A | N/A |
| 21 | - | wan2.6-t2v | Alibaba | 1327 ±7 | 54050 | N/A | N/A |
| 22 | - | seedance-v1.5-pro | Bytedance | 1256 ±7 | 76171 | N/A | N/A |
| 25 | - | wan2.5-t2v-preview | Alibaba | 1245 ±9 | 36378 | N/A | N/A |
本周 Arena AI 大模型周榜总体头部格局变化不大,海外模型依旧占据榜首位置,但国产模型在多个细分榜单已经稳定站稳中上游,kimi-k3-max 在代码榜进入前十,mimo-v2.6-pro 在前端开发榜和智能体榜均有亮眼表现,字节跳动、阿里的国产模型在 AI 生图、AI 视频领域也保持了前五、前十的好成绩。后续笔者将持续关注国产大模型的进展,带来最新排名变化。
本文来源:https://www.ithome.com/1/014/865.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT House October 19 news, Arena (arena.ai) platform released the 42nd week of 2026 (2026-10-12 ~ 2026-10-18) AI big model blind test ranking, the current head pattern is generally stable, the domestic model in a number of subdivisions of the list to maintain a stable upstream position, of which kimi-k3-max for the first time stable in the top ten of the code list, mimo-v2.6-pro for the first time into the top 20 front-end development list, the overall performance is remarkable. There are three new models officially listed in this issue, namely gemini-3.5-flash-medium, mimo-v2.6-pro and flux-2-flex.
IT House Note: This list is based on the anonymous blind polling of the Arena (arena.ai) platform. The ranking is calculated by using the ELO (agent list is a percentage signal) to reflect the user's subjective preferences rather than the absolute ability test. Different rankings are independent of each other and should not be compared across the list. The score difference is considered to be juxtaposed within the error range. The new model also needs to accumulate a certain number of votes before it will be officially included in the list.
The top ranking of this comprehensive list remained stable, with gemini-4-argon-high returning to the top with 1525 points, claude-opus-5.5-high ranking second with 1507 points, claude-opus-4-6-high and claude-fable-5-high tied for third with 1504 points, and the scores were all within 30 points, which was regarded as close to competition within the margin of error. The scores in the top 10 of the head are all above 1494, and the competition is still stuck. This week's new entry model is gemini-3.5-flash-medium, directly rushing to the top 30, ranking 30th, with an ELO score of 1476, which is a good performance.
The domestic model is generally in the middle and upstream position, and the echelon is relatively stable:
The header pattern of the code list is still stable, gemini-4-argon-high continues to lead with 1562 points, claude-fable-5-high ranks second with 1551 points, claude-opus-4-7-high, claude-opus-4-6-high, claude-opus-5.5-high ties with 1550 points, the top three are all Anthropic models except gemini-4-argon-high, and the head is still occupied by overseas manufacturers. The core point of view this week is that the domestic kimi-k3-max remained stable at 10th place, with an ELO of 1540, becoming the highest ranking of domestic models in the code list, followed by the millet mimo-v2.6-pro at 11th place, with an ELO of 1539, showing brilliant performance.
Various domestic models have entered the top 30 of the code list, and the specific ranking is as follows:
The top pattern of the front-end development list is stable, with claude-opus-5.5-max at the top with a score of 1813, gpt-6-astra-max at the second place with a score of 1786, and claude-sonnet-5.5-xhigh at the third place with a score of 1774. This week's new entry model covers mimo-v2.6-pro and claude-haiku-5.5-high, of which Xiaomi mimo-v2.6-pro first entered the list and ranked 19th, ELO 1629 points, successfully entering the top 20, continuing its good performance in the code field.
Domestic models occupy multiple mid- and upstream positions in the front-end development list, as follows:

The agent list uses a percentage signal scoring system. This week, Claude Opus 5.5 (Max) ranked first with a net improvement of 15.53%, GPT 6 Astra (Max) ranked second with a net improvement of 13.09%, and Claude Fable 5.1 (Max) ranked third with a net improvement of 12.66%. The head is still occupied by overseas models. In the indicator dimension, the Claude Opus 5.5 (Max) tool hallucination rate is only 0.37%, which is the lowest on the head, and it performs well. Domestic models are generally in the middle and upstream echelons, and many have entered the top 30.
The specific performance of domestic models in the agent list is as follows:
The top pattern of the AI chart is stable, with gpt-image-2.5-sunburst ranking first with 1425 points, gpt-image-2.5-flare ranking second with 1398 points, gpt-image-2 (medium) ranking third with 1383 points, and the top three are all OpenAI models. This week's new entry model is flux-2-flex, which ranks 30th when it first enters the list, with an ELO score of 1157 points. The domestic model performed steadily in the list, with ByteDance seedream-5.0-pro and Ali qwen-image-3.0-pro tied for 12th and 13th with 1255 points, both in the top 15, in the middle and upstream positions.
The top of the AI video chart is still dominated by Google models, with gemini-omni-1.1-flash in first place with a score of 1516, gemini-omni-flash in second place with a score of 1513, and flux-3-video in third place with a score of 1493. Domestic models performed prominently, occupying two seats in the top five, byte beating dreamina-seedance-2.0-720p ranked fifth, ELO 1479 points, Ali wan3.0 followed by sixth, ELO 1476 points, there are four domestic models in the top 10, and the overall strength should not be underestimated.
This week, the overall headline pattern of the Arena AI Weekly Model List has not changed much, and overseas models still occupy the top position. However, domestic models have stabilized in the middle and upstream of many subdivision lists, kimi-k3-max has entered the top ten in the code list, mimo-v2.6-pro has performed brilliantly in the front-end development list and the agent list, and ByteDance, Ali's domestic models have also maintained top five and top ten in the field of AI raw graphics and AI video. In the future, I will continue to pay attention to the progress of domestic large models, bringing the latest ranking changes.
This article is from: https://www.ithome.com/1/014/865.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.