OfficePrism
Agents resolve Chinese workplace tickets using one business system at a time. The evaluator checks outcomes across 420 tasks in seven business functions.
- Made by
- Alibaba Group and collaborators
- Released
- November 2026
- Kind
- Documents and spreadsheets
- Level
- University
- Scored
- Graded by a model
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Subject
- Economics and business
- Language
- Chinese
- Interaction
- Agentic
- Environment
- APIs and tools
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Agents, Law and finance
- Strongest model tested
- GPT-5.5 #21 on the capability index
- Last model added
- July 2026
- Scores refreshed
- November 20th, 2026
Kimi K3 leads with 76.4% among models released by July 2026
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
The newest model with a score here came out in July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 Kimi K3 full 76.4%
- 2 GLM-5.2 search 74.3%
- 3 Claude Opus 4.7 search 74%
- 4 Claude Opus 4.8 search 74%
- 5
DeepSeek-V4-Flash full 73.6%
- 6 Claude Sonnet 5 search 73.1%
- 7
DeepSeek-V4-Pro full 72.9%
- 8
Qwen3.7-Plus full 72.9%
- 9
Qwen3.7-Max search 72.6%
- 10 GLM-5.1 search 72.4%
- 11 GPT-5.5 search 71.9%
- 12
Qwen 3.6 Max (Preview) search 71.2%
- 13
Qwen 3.6 Plus full 69.5%
- 14 Kimi K2.6 full 66%
Source: The OfficePrism paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Kimi K3 | 76.4% | full, ReAct | Jul 2026 | Official leaderboard |
| 2 | GLM-5.2 | 74.3% | search, Claude Code | Jun 2026 | Official leaderboard |
| 3 | Claude Opus 4.7 | 74% | search, ReAct | Apr 2026 | Official leaderboard |
| 4 | Claude Opus 4.8 | 74% | search, ReAct | May 2026 | Official leaderboard |
| 5 | | 73.6% | full, ReAct | Apr 2026 | Official leaderboard |
| 6 | Claude Sonnet 5 | 73.1% | search, ReAct | Jun 2026 | Official leaderboard |
| 7 | | 72.9% | full, ReAct | Apr 2026 | Official leaderboard |
| 8 | | 72.9% | full, Claude Code | Jun 2026 | Official leaderboard |
| 9 | | 72.6% | search, Claude Code | May 2026 | Official leaderboard |
| 10 | GLM-5.1 | 72.4% | search, Claude Code | Apr 2026 | Official leaderboard |
| 11 | GPT-5.5 | 71.9% | search, DeepAgents | Apr 2026 | Official leaderboard |
| 12 | | 71.2% | search, Claude Code | Apr 2026 | Official leaderboard |
| 13 | | 69.5% | full, Claude Code | Mar 2026 | Official leaderboard |
| 14 | Kimi K2.6 | 66% | full, ReAct | Apr 2026 | Official leaderboard |
Rankings by part
We read 6 parts from the benchmark’s paper, each ranking models on one part of the test, such as L1, L2 and L3. Kimi K3 leads 4 of them, and Claude Sonnet 5 leads the other 2. Arrows show how far a model moved from its overall rank among the models we track.
L1
14 models we track have a score here.
- 1 Claude Sonnet 5 ↑5 93.3%
- 2 Claude Opus 4.7 ↑1 92.4%
- 3 Claude Opus 4.8 ↑1 92.4%
- 4 Kimi K3 ↓3 91.4%
- 5
Qwen3.7-Plus ↑3 90.5%
- 6
Qwen3.7-Max ↑3 89.5%
- 7 GLM-5.1 ↑3 88.6%
- 8 Kimi K2.6 ↑6 88.6%
- 9 GLM-5.2 ↓7 88.6%
- 10
Qwen 3.6 Plus ↑3 87.6%
L2
14 models we track have a score here.
- 1 Kimi K3 82.9%
- 2 GLM-5.2 82.1%
- 3 GLM-5.1 ↑7 81.4%
- 4
DeepSeek-V4-Pro ↑3 80.7%
- 5
Qwen3.7-Max ↑4 80%
- 6 Claude Sonnet 5 79.3%
- 7 GPT-5.5 ↑4 78.6%
- 8
DeepSeek-V4-Flash ↓3 78.6%
- 9 Kimi K2.6 ↑5 77.9%
- 10
Qwen3.7-Plus ↓2 77.1%
L3
14 models we track have a score here.
- 1 Kimi K3 65.7%
- 2 Claude Opus 4.7 ↑1 64.6%
- 3
Qwen 3.6 Max (Preview) ↑9 64.6%
- 4 Claude Opus 4.8 64%
- 5
DeepSeek-V4-Flash 62.9%
- 6 GLM-5.2 ↓4 62.9%
- 7
Qwen3.7-Plus ↑1 61.7%
- 8
Qwen 3.6 Plus ↑5 61.1%
- 9
DeepSeek-V4-Pro ↓2 60%
- 10 GLM-5.1 59.4%
Success
14 models we track have a score here.
- 1 Kimi K3 85.5%
- 2 GLM-5.2 82.4%
- 3
DeepSeek-V4-Flash ↑2 81.4%
- 4
DeepSeek-V4-Pro ↑3 81%
- 5 Claude Opus 4.8 ↓1 81%
- 6 Claude Opus 4.7 ↓3 80.5%
- 7 GLM-5.1 ↑3 80.2%
- 8
Qwen3.7-Plus 80.2%
- 9 Claude Sonnet 5 ↓3 80.2%
- 10
Qwen3.7-Max ↓1 80%
Strict
14 models we track have a score here.
- 1 Kimi K3 61.4%
- 2 GLM-5.2 61.2%
- 3 GLM-5.1 ↑7 60%
- 4
Qwen3.7-Max ↑5 59.8%
- 5
DeepSeek-V4-Pro ↑2 59.5%
- 6
Qwen3.7-Plus ↑2 59.5%
- 7
DeepSeek-V4-Flash ↓2 59.3%
- 8 Claude Sonnet 5 ↓2 58.6%
- 9 GPT-5.5 ↑2 58.1%
- 10 Claude Opus 4.8 ↓6 58.1%
Mean recorded steps
14 models we track have a score here.
- 1 Claude Sonnet 5 ↑5 4.80
- 2
Qwen 3.6 Max (Preview) ↑10 6.40
- 3 Claude Opus 4.7 6.70
- 4
Qwen3.7-Max ↑5 6.70
- 5
Qwen3.7-Plus ↑3 6.80
- 6 Claude Opus 4.8 ↓2 7.00
- 7
Qwen 3.6 Plus ↑6 7.20
- 8 GLM-5.1 ↑2 7.30
- 9 Kimi K2.6 ↑5 7.40
- 10 Kimi K3 ↓9 7.50
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The OfficePrism paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Kimi K3 | full | ReAct | 76.4% | ||
| Kimi K3 | search | ReAct | 76.2% | ||
| Kimi K3 | full | DeepAgents | 76% | ||
| Kimi K3 | search | DeepAgents | 76% | ||
| GLM-5.2 | search | Claude Code | 74.3% | ||
| Claude Opus 4.7 | search | ReAct | 74% | ||
| Claude Opus 4.8 | search | ReAct | 74% | ||
| full | ReAct | 73.6% | |||
| Claude Sonnet 5 | search | ReAct | 73.1% | ||
| full | ReAct | 72.9% | |||
| full | Claude Code | 72.9% | |||
| Kimi K3 | search | Claude Code | 72.9% | ||
| Claude Opus 4.7 | search | DeepAgents | 72.6% | ||
| search | Claude Code | 72.6% | |||
| search | Claude Code | 72.6% | |||
| GLM-5.2 | search | Hermes | 72.6% | ||
| Claude Sonnet 5 | search | DeepAgents | 72.6% | ||
| GLM-5.1 | search | Claude Code | 72.4% | ||
| Claude Opus 4.8 | search | Claude Code | 72.4% | ||
| search | ReAct | 72.1% | |||
| search | Claude Code | 72.1% | |||
| GLM-5.2 | full | ReAct | 72.1% | ||
| Kimi K3 | search | OpenClaw | 72.1% | ||
| Claude Opus 4.7 | search | Claude Code | 71.9% | ||
| GPT-5.5 | search | DeepAgents | 71.9% | ||
| GLM-5.2 | search | ReAct | 71.9% | ||
| Kimi K3 | search | Hermes | 71.9% | ||
| GLM-5.1 | full | Claude Code | 71.7% | ||
| GLM-5.1 | search | ReAct | 71.4% | ||
| GLM-5.2 | full | Claude Code | 71.4% | ||
| search | Claude Code | 71.2% | |||
| Claude Opus 4.8 | search | DeepAgents | 71.2% | ||
| Claude Sonnet 5 | full | DeepAgents | 71.2% | ||
| GPT-5.5 | search | ReAct | 71% | ||
| search | ReAct | 71% | |||
| search | DeepAgents | 71% | |||
| GLM-5.2 | search | DeepAgents | 70.7% | ||
| search | DeepAgents | 70.2% | |||
| GLM-5.2 | full | DeepAgents | 70.2% | ||
| GLM-5.2 | full | Hermes | 70.2% | ||
| Kimi K3 | full | Claude Code | 70.2% | ||
| GLM-5.1 | full | ReAct | 70% | ||
| Claude Opus 4.7 | full | DeepAgents | 70% | ||
| Claude Opus 4.7 | search | Hermes | 70% | ||
| Claude Opus 4.8 | full | DeepAgents | 70% | ||
| search | ReAct | 70% | |||
| full | Claude Code | 69.5% | |||
| search | Claude Code | 69.5% | |||
| Claude Opus 4.8 | search | Hermes | 69.5% | ||
| full | ReAct | 69.5% | |||
| Kimi K3 | full | Hermes | 69.3% | ||
| search | Claude Code | 69% | |||
| GLM-5.1 | search | DeepAgents | 69% | ||
| search | OpenClaw | 69% | |||
| Claude Sonnet 5 | search | Hermes | 69% | ||
| Claude Sonnet 5 | search | Claude Code | 68.8% | ||
| full | Claude Code | 68.6% | |||
| full | Hermes | 68.3% | |||
| GPT-5.5 | full | DeepAgents | 68.1% | ||
| search | DeepAgents | 68.1% | |||
| search | Hermes | 68.1% | |||
| full | DeepAgents | 67.9% | |||
| full | Hermes | 67.9% | |||
| full | DeepAgents | 67.9% | |||
| full | DeepAgents | 67.9% | |||
| GLM-5.2 | full | OpenClaw | 67.9% | ||
| full | ReAct | 67.6% | |||
| full | Claude Code | 67.6% | |||
| full | Claude Code | 67.6% | |||
| GLM-5.2 | search | OpenClaw | 67.6% | ||
| search | DeepAgents | 67.4% | |||
| GLM-5.1 | full | DeepAgents | 67.1% | ||
| full | ReAct | 66.9% | |||
| search | Hermes | 66.9% | |||
| Kimi K3 | full | OpenClaw | 66.7% | ||
| search | Hermes | 66.4% | |||
| full | DeepAgents | 66.2% | |||
| full | DeepAgents | 66.2% | |||
| Kimi K2.6 | full | ReAct | 66% | ||
| full | OpenClaw | 65.7% | |||
| GLM-5.1 | search | OpenClaw | 65.5% | ||
| search | OpenClaw | 65.5% | |||
| full | ReAct | 65.5% | |||
| search | OpenClaw | 65.2% | |||
| Kimi K2.6 | full | DeepAgents | 65% | ||
| search | ReAct | 65% | |||
| search | OpenClaw | 64.5% | |||
| GLM-5.1 | full | OpenClaw | 64.3% | ||
| full | Hermes | 64.3% | |||
| GLM-5.1 | search | Hermes | 64% | ||
| search | ReAct | 63.8% | |||
| full | DeepAgents | 63.6% | |||
| Claude Opus 4.7 | search | OpenClaw | 63.3% | ||
| Kimi K2.6 | search | ReAct | 63.3% | ||
| search | ReAct | 63.3% | |||
| full | Claude Code | 63.1% | |||
| GLM-5.1 | full | Hermes | 62.9% | ||
| Kimi K2.6 | search | DeepAgents | 62.9% | ||
| search | DeepAgents | 62.9% | |||
| search | OpenClaw | 62.9% | |||
| search | Hermes | 62.6% | |||
| Kimi K2.6 | search | Claude Code | 62.4% | ||
| GPT-5.5 | search | Claude Code | 62.4% | ||
| search | Hermes | 62.4% | |||
| search | OpenClaw | 62.1% | |||
| full | OpenClaw | 61.7% | |||
| Kimi K2.6 | search | Hermes | 61.4% | ||
| full | OpenClaw | 61.4% | |||
| full | Hermes | 61.4% | |||
| Claude Opus 4.8 | search | OpenClaw | 61.4% | ||
| Kimi K2.6 | full | Claude Code | 61.2% | ||
| full | Hermes | 61% | |||
| GPT-5.5 | search | Hermes | 60.5% | ||
| Kimi K2.6 | search | OpenClaw | 59.8% | ||
| full | OpenClaw | 59.8% | |||
| full | Hermes | 59.5% | |||
| full | OpenClaw | 59.3% | |||
| Kimi K2.6 | full | Hermes | 59.3% | ||
| search | DeepAgents | 58.3% | |||
| Kimi K2.6 | full | OpenClaw | 58.3% | ||
| full | OpenClaw | 54% | |||
| search | Hermes | 53.6% | |||
| GPT-5.5 | search | OpenClaw | 32.9% | ||
| Claude Sonnet 5 | search | OpenClaw | 22.6% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The OfficePrism paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official OfficePrism leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Economic work
- JobBench In use 130 tasks from 35 occupations workers most want delegated. Agent works through messy reference files, graded by fact-anchored binary rubrics. GLM-5.3 61.4%
- GDPval In use Real work tasks from 44 occupations with reference files. Experts blindly compare model deliverables with expert work. Metric is win-or-tie rate.
Seed2.1 Pro 87.9%
- GDPval-AA v2.1 In use Agents with shell and web access complete 220 GDPval gold tasks from 44 occupations. Blind pairwise comparisons give Elo ratings. Claude Opus 5.5 1,846
- AA-Briefcase v1.1 In use Version 1.1 of Artificial Analysis's multi-hour office work evaluation, with Elo ratings fitted by Crowd-BT. Claude Opus 5.5 1,822
- Agents' Last Exam v1.1 In use Agents complete 122 public long-horizon professional workflow tasks across 55 sub-industries. Deterministic graders check outcomes against hidden references. GPT-6.2 Astra 47.5%
- CoWorkBench In use Agent completes long-horizon productivity tasks across computer science, finance, law and medicine. Qwen in-house scoring (method unpublished). Claude Fable 5 75.9%
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (14 entries).
- Pass Models scored 14 models we track.
- Pass Newest score The newest is from July 2026, for Kimi K3.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
- Pass Headroom The best is 76.4%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 124 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.