Research PreviewLeaderboard v1.2 · 2026-10-15

Tencent WorkBuddy Bench Tencent WorkBuddy Bench

Can coding agents solve real repository, front-end and office tasks end-to-end? 编程 Agent 能否端到端解决真实的仓库、前端与办公任务?

Loading leaderboard…

01

Key Takeaways核心结论

  • Claude Opus 5.5 sweeps the core tracks: under CodeBuddy Code it leads Code at 83.26, Web at 76.24, and Office at 85.24, while DeepSeek-V4.1-Flash-1017 leads Security at 72.22. Under Claude Code, GLM-5.2 tops Code, and HY-3 leads Web, Office and Security.Claude Opus 5.5 横扫核心赛道：CodeBuddy Code 下以 83.26 / 76.24 / 85.24 领跑 Code、Web、Office，DeepSeek-V4.1-Flash-1017 以 72.22 领跑 Security；Claude Code 下 GLM-5.2 拿下 Code，HY-3 领跑 Web、Office 与 Security。
  • Open-weight models hold their own: they lead Security and every Claude Code column, with GLM-5.2 topping Code and HY-3 topping Web, Office and Security under Claude Code; open-weight competitiveness here is board-dependent, not uniform.开源权重模型不落下风:Security 及 Claude Code 下各列榜首均为开源模型 —— GLM-5.2 拿下 Claude Code 下的 Code 榜首,HY-3 领跑 Web、Office 与 Security。开源权重的竞争力在这套基准上因榜单而异,并非全面领先。
  • The harness is part of the result: the same model can swing hard between harnesses — GLM-5.2 on Web scores 67.43 vs 60.71, DeepSeek-V4-Flash on Security 65.65 vs 53.90 — so read scores per harness.harness 是结果的一部分:同一模型在不同 harness 下可能大幅波动 —— GLM-5.2 在 Web 上 67.43 vs 60.71,DeepSeek-V4-Flash 在 Security 上 65.65 vs 53.90，分数应按 harness 分开解读。
  • Efficiency does not follow rank: the leanest output budgets sit at both the top and the middle of the board, and some mid-table runs spend 3–4× the tokens of leaner models at similar scores.效率与排名并不一致:输出 token 预算最小的模型既有榜首也有中游,一些中游成绩却要花 3–4 倍 的 token。
  • Integration details move scores: enabling cross-turn reasoning passback for HY-3 lifted its Code score by +3.82 (CodeBuddy Code) / +1.92 (Claude Code) in a diagnostic rerun.接入细节会左右分数:在一次诊断性复跑中,为 HY-3 开启跨轮推理回传后,其 Code 分数提升 +3.82(CodeBuddy Code)/ +1.92(Claude Code)。
02

Token efficiencyToken 效率

score vs output tokens per run, 3-run avg分数 vs 每次运行输出 token，三次运行平均

CodeBuddy Code 2.109.3 · Claude Code 2.1.187 · think

Efficiency does not follow rank order: GLM-5.2’s 77.06 on Code under Claude Code costs 22.0k output tokens per run, while leaner mid-table models reach similar scores on a fraction of that budget. Security is by far the heaviest subset: mid-board runs there average up to roughly 90 turns and around 11M cache-inclusive input tokens per run under CodeBuddy Code. 效率与排名并不一致：GLM-5.2 在 Claude Code 下 Code 的 77.06 每次运行花费 22.0k 输出 token，而更精简的中游模型只花其中一小部分预算就能拿到相近分数。Security 是 token 消耗最大的子集：中游成绩在 CodeBuddy Code 下平均每次运行可达约 90 轮交互、约 1,100 万输入 token（含缓存）。

All points are 3-run averages in think mode. Turns count unique assistant messages including subagent activity; output tokens come from each run’s final usage. Output tokens are the comparable axis — input-token accounting differs across harness configurations, so input tokens are never compared across harnesses. 所有数据点均为 3 次运行平均、think 模式。轮数按去重后的 assistant 消息计数（含 subagent 活动）；输出 token 取自每次运行的最终 usage。图中仅比较输出 token，输入 token 的统计口径随 harness 配置不同，因此不做跨 harness 的输入 token 比较。