Head-to-head in the coding agent arena: Opus 4.8 holds a slight edge on NL2Repo and Toolathlon, while V4-Pro pulls ahead on CyberGym and HLE, with the price gap holding steady at an order of magnitude.
| Benchmark | GLM-5.3 | Kimi K3 | DeepSeek-V4-Pro | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.2 | 88.3 | 87.9 | 85.0 | 88.8 |
| Terminal-Bench 3.0 | 28.3 | 17.4 | — | 21.1 | 34.6 |
| DeepSWE v1.1 | 66.9 | 67.5 | 62.7 | 58.0 | 72.7 |
| NL2Repo | 58.0 | 58.0 | 61.1 | 69.7 | — |
| CyberGym | 84.5 | 80.0 | 83.3 | 78.1 | 83.6 |
| Toolathlon Verified | 73.0 | 76.5 | 74.1 | 76.2 | 74.9 |
| AutomationBench v1.0.6 | 48.2 | 46.7 | 43.2 | 41.0 | 45.8 |
| ALE-CLI | 28.5 | 27.6 | 25.7 | 25.7 | 28.6 |
| HLE w/ Tools | 62.5 | 59.8 | 60.0 | 57.9 | 64.5 |
| GDPval-AA v2 | 1769 | 1682 | 1590 | 1588 | 1730 |
Data source: Z.ai official blog (2026-08-14); not evaluated is indicated by —. DeepSeek pricing: V4-Pro off-peak $0.66/$1.98, V4-Flash $0.14/$0.28, V4.1-Flash-1017 $0.15/$0.60 per million tokens.
Opus 4.8: NL2Repo 69.7, Toolathlon 76.2, AutomationBench 41.0; V4-Pro: 61.1 / 74.1 / 43.2 — they trade wins and losses, with gaps in the single-digit percentage points.
Both offer 1M-class context windows; third-party evaluations note that Opus 4.6/4.8's MRCR retrieval recall still leads, while DeepSeek leverages CSA+HCA to push costs down to 27% FLOPs / 10% KV.
CyberGym: V4-Pro 83.3 vs Opus 78.1—DeepSeek currently leads in the vulnerability discovery track; on ExploitGym, GPT/Fable are more aggressive.
If you want the best coding experience among closed-source models and long-context retrieval precision, choose Claude; if you're budget-sensitive, need security auditing capabilities, or want self-hosting, choose DeepSeek.
大模型网摘