Key findings
- Kimi K3 is the best Chinese model in every one of the 12 fields, and reaches the top ten in consulting, legal and finance.
- Kimi K3 trails the overall leader by 188 points, but GPT-6.1 Sol on OpenCode by only 49.
- DeepSeek Flash on Codex is on the quality–cost frontier at $0.11 a task, and in medical tasks it is both the cheapest and the fastest pairing on the board.
- GLM-5.3 on OpenCode scores 28 points below GLM-5.3 Prime's best pairing at less than a fifth of the cost.
Where each model stands
The table takes each model's best pairing on the overall board.
| Overall place | Model | Best harness | Quality | Cost per task | Time per run |
|---|---|---|---|---|---|
| 13 | Kimi K3 | OpenCode | 1147 | $1.23 | 904.6s |
| 19 | DeepSeek Flash | Claude Code | 1039 | $0.18 | 1050.7s |
| 22 | GLM-5.3 Prime | Claude Code | 1021 | $2.51 | 977.1s |
| 23 | Qwen3.8 Max | Codex | 1001 | $1.30 | 1800.6s |
| 27 | GLM-5.3 | OpenCode | 993 | $0.47 | 710.5s |
| 32 | DeepSeek V4 Pro | Claude Code | 914 | $0.41 | 799.1s |
| 38 | Qwen3.8 Flash | Codex | 859 | $0.15 | 1562.3s |
| 39 | Qwen3.7 Plus | Codex | 856 | $0.15 | 586.0s |
| 44 | MiniMax M2.5 | OpenCode | 700 | $0.17 | 751.7s |
| 48 | Kimi K2.7 Code | Kimi CLI | 651 | $0.30 | 464.3s |
Four of the ten score above the 1000 reference. For comparison, the overall leader, Claude Opus 5.5 on Claude Code, scores 1335, and the median of all 49 pairings is 999.
Kimi K3
Kimi K3 ran on four harnesses, and all four pairings place in the overall top twenty: OpenCode 1147, Codex 1110, Claude Code 1095 and Moonshot's own Kimi CLI 1094.
By field, its best places are 8th in consulting, 9th in legal and 10th in finance. In the other nine fields it places between 11th and 15th, and in each of them it is still the highest-placed Chinese model.
Against the leaders, 188 points behind Claude Opus 5.5 means an expected win rate of about 25% against it on the same task. The gap to the GPT-6 Sol models is much smaller: 49 points to GPT-6.1 Sol on OpenCode and 43 to GPT-6 Sol on OpenCode. Kimi K3 costs more than either, $1.23 a task against about $0.50.
DeepSeek
DeepSeek Flash is the best value among the Chinese models. On Codex it scores 1031 for $0.11 a task, which puts it on the overall quality–cost frontier: no pairing scores higher for less. On Claude Code it scores 1039 for $0.18.
It is strongest relative to its price in medical tasks. There DeepSeek Flash on Codex is the cheapest pairing on the board at $0.004 a task and also the fastest at 81.6 seconds a run, while scoring 1038. In 3D, where the leader spends $17.73 a task, DeepSeek Flash on Claude Code scores 1105 for $0.16.
DeepSeek V4 Pro places lower, 914 at best. On DeepSeek's own harness it scores 893 but is its cheapest and fastest setup, at $0.18 a task and 488.5 seconds a run.
GLM Qwen and MiniMax
- GLM-5.3 Prime scores 1021 on Claude Code but costs $2.51 a task, close to Claude Opus 5.5's $2.60. GLM-5.3 on OpenCode scores 993 for $0.47. For most uses the standard model is the better buy.
- Qwen3.8 Max scores between 998 and 1001 on all three harnesses it ran, but it is slow: 1253.1 to 1800.6 seconds a run, including the slowest pairing on the board.
- Qwen3.8 Flash on OpenCode is the cheapest pairing on the board at $0.049 a task, scoring 832. On Codex it scores 859 for $0.15.
- MiniMax M2.5 scores 700 at best, on OpenCode, and Kimi K2.7 Code 651 on Kimi CLI.
The harness matters more for these models
For several Chinese models the choice of harness moved quality more than for the leading models: 61 points for GLM-5.3, 59 for MiniMax M2.5, 53 for Kimi K3 and 39 for DeepSeek Flash. All three Chinese models that ran on both their vendor's own harness and general-purpose ones scored higher on at least one general-purpose harness. The harness analysis covers this across all models.
Questions
What is the best Chinese AI model for agent work
On the November 2026 Work Arena leaderboard, Kimi K3 is the highest-placed Chinese model, 13th overall at 1147 on OpenCode, and the top Chinese model in all 12 fields.
Is DeepSeek good for AI agent tasks
DeepSeek Flash is a strong low-cost option. On Codex it scores 1031, above the 1000 reference, for $0.11 a task, and no pairing on the board scores higher for less. DeepSeek V4 Pro scores lower, 914 at best.
Figures in this article come from the leaderboard data of 2 November 2026. The leaderboard shows any changes since.