OdysseyBench
Agents recover requirements from conversation histories and complete workflows across documents, spreadsheets, PDFs, email, and calendars. Task-specific checks inspect the resulting file system and report the percentage of completed tasks.
OdysseyBench+ contains 300 tasks and OdysseyBench-Neo contains 302 synthesized tasks. Long-context and retrieval settings use different input histories and should remain separate. The suite builds on OfficeBench.
- Made by
- Microsoft and University of Edinburgh
- Released
- August 2025
- Kind
- Documents and spreadsheets
- Level
- University
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Desktop apps, APIs and tools
- Items from
- Another benchmark, Generated
- Refresh
- One fixed set
- Run by
- Anyone
- Based on
- A version of another benchmark
- In views
- Agents
- Strongest model tested
- GPT-5 #71 on the capability index
- Last model added
- August 2025
- Scores refreshed
- November 20th, 2026
In the authors’ words
Autonomous agents powered by large language models (LLMs) are increasingly deployed in real-world applications requiring complex, long-horizon workflows. However, existing benchmarks predominantly focus on atomic tasks that are self-contained and independent, failing to capture the long-term contextual dependencies and multi-interaction coordination required in realistic scenarios. To address this gap, we introduce OdysseyBench, a comprehensive benchmark for evaluating LLM agents on long-horizon workflows across diverse office applications including Word, Excel, PDF, Email, and Calendar. Our benchmark comprises two complementary splits: OdysseyBench+ with 300 tasks derived from real-world use cases, and OdysseyBench-Neo with 302 newly synthesized complex tasks. Each task requires agent to identify essential information from long-horizon interaction histories and perform multi-step reasoning across various applications. To enable scalable benchmark creation, we propose HomerAgents, a multi-agent framework that automates the generation of long-horizon workflow benchmarks through systematic environment exploration, task generation, and dialogue synthesis. Our extensive evaluation demonstrates that OdysseyBench effectively challenges state-of-the-art LLM agents, providing more accurate assessment of their capabilities in complex, real-world contexts compared to existing atomic task benchmarks. We believe that OdysseyBench will serve as a valuable resource for advancing the development and evaluation of LLM agents in real-world productivity scenarios. In addition, we release OdysseyBench and HomerAgents to foster research along this line.
o3 leads with 58.7% among models released by August 2025 o3 leads with 58.7% among models released by August 2025
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The paper reports no overall success rate, so this is the plain mean of the OdysseyBench+ and OdysseyBench-Neo overall percentages. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in August 2025, 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 o3 Long-context prompting 58.7%
- 2 GPT-5 Long-context prompting 55%
- 3 GPT-5 Chat Long-context prompting 49%
- 4 DeepSeek-R1 (snapshot unspecified) Long-context prompting 47.1%
- 5 GPT-4.1 Long-context prompting 46.1%
- 6 GPT-4o (snapshot unspecified) Retrieval: summary chunk, top 25 44.8%
- 7 o3-mini Long-context prompting 36.5%
- 8 GPT-4o-mini (snapshot unspecified) Long-context prompting 28.4%
- 9
Qwen3-32B Long-context prompting 26.3%
- 10
DeepSeek-R1-Distill-Qwen-32B Long-context prompting 17.3%
Source: The OdysseyBench paper
- 1 o3 Long-context prompting 58.7%
- 2 GPT-5 Long-context prompting 55%
- 3 GPT-5 Chat Long-context prompting 49%
- 4 GPT-4.1 Long-context prompting 46.1%
- 5 o3-mini Long-context prompting 36.5%
- 6
Qwen3-32B Long-context prompting 26.3%
- 7
DeepSeek-R1-Distill-Qwen-32B Long-context prompting 17.3%
Source: The OdysseyBench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | o3 | 58.7% | Long-context prompting | Apr 2025 | Official leaderboard |
| 2 | GPT-5 | 55% | Long-context prompting | Aug 2025 | Official leaderboard |
| 3 | GPT-5 Chat | 49% | Long-context prompting | Aug 2025 | Official leaderboard |
| 4 | DeepSeek-R1 (snapshot unspecified) | 47.1% | Long-context prompting | Official leaderboard | |
| 5 | GPT-4.1 | 46.1% | Long-context prompting | Apr 2025 | Official leaderboard |
| 6 | GPT-4o (snapshot unspecified) | 44.8% | Retrieval: summary chunk, top 25 | Official leaderboard | |
| 7 | o3-mini | 36.5% | Long-context prompting | Jan 2025 | Official leaderboard |
| 8 | GPT-4o-mini (snapshot unspecified) | 28.4% | Long-context prompting | Official leaderboard | |
| 9 | | 26.3% | Long-context prompting | Apr 2025 | Official leaderboard |
| 10 | | 17.3% | Long-context prompting | Jan 2025 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | o3 | 58.7% | Long-context prompting | Apr 2025 | Official leaderboard |
| 2 | GPT-5 | 55% | Long-context prompting | Aug 2025 | Official leaderboard |
| 3 | GPT-5 Chat | 49% | Long-context prompting | Aug 2025 | Official leaderboard |
| 4 | GPT-4.1 | 46.1% | Long-context prompting | Apr 2025 | Official leaderboard |
| 5 | o3-mini | 36.5% | Long-context prompting | Jan 2025 | Official leaderboard |
| 6 | | 26.3% | Long-context prompting | Apr 2025 | Official leaderboard |
| 7 | | 17.3% | Long-context prompting | Jan 2025 | Official leaderboard |
Rankings by part
We read 8 parts from the benchmark’s paper, each ranking models on one part of the test, such as OdysseyBench+ 1-apps, OdysseyBench+ 2-apps and OdysseyBench+ 3-apps. o3 leads 5 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
OdysseyBench+ 1-apps
7 models we track have a score here.
- 1 GPT-5 ↑1 75.3%
- 2 o3 ↓1 72.8%
- 3 GPT-4.1 ↑1 55.9%
- 4 GPT-5 Chat ↓1 55.9%
- 5
Qwen3-32B ↑1 38.7%
- 6 o3-mini ↓1 38%
- 7
DeepSeek-R1-Distill-Qwen-32B 30.1%
OdysseyBench+ 2-apps
7 models we track have a score here.
- 1 o3 70.5%
- 2 GPT-5 66.3%
- 3 GPT-5 Chat 48.4%
- 4 GPT-4.1 43.2%
- 5
Qwen3-32B ↑1 33.7%
- 6 o3-mini ↓1 20%
- 7
DeepSeek-R1-Distill-Qwen-32B 16.8%
OdysseyBench+ 3-apps
7 models we track have a score here.
- 1 o3 30.4%
- 2 GPT-5 25.9%
- 3 GPT-5 Chat 20.5%
- 4 o3-mini ↑1 15.2%
- 5 GPT-4.1 ↓1 12.5%
- 6
Qwen3-32B 11.6%
- 7
DeepSeek-R1-Distill-Qwen-32B 1.79%
OdysseyBench+ overall
7 models we track have a score here.
- 1 o3 56.2%
- 2 GPT-5 54%
- 3 GPT-5 Chat 40.3%
- 4 GPT-4.1 35.7%
- 5
Qwen3-32B ↑1 27%
- 6 o3-mini ↓1 23.8%
- 7
DeepSeek-R1-Distill-Qwen-32B 15.3%
OdysseyBench-Neo 1-apps
7 models we track have a score here.
- 1 GPT-4.1 ↑3 75%
- 2 GPT-5 Chat ↑1 75%
- 3 o3-mini ↑2 71.7%
- 4 o3 ↓3 68.3%
- 5 GPT-5 ↓3 61.7%
- 6
Qwen3-32B 41.7%
- 7
DeepSeek-R1-Distill-Qwen-32B 40%
OdysseyBench-Neo 2-apps
7 models we track have a score here.
- 1 GPT-4.1 ↑3 63.4%
- 2 o3 ↓1 60.6%
- 3 GPT-5 Chat 57.8%
- 4 GPT-5 ↓2 56.3%
- 5 o3-mini 39.4%
- 6
DeepSeek-R1-Distill-Qwen-32B ↑1 22.5%
- 7
Qwen3-32B ↓1 22.5%
OdysseyBench-Neo 3-apps
7 models we track have a score here.
- 1 o3 59.1%
- 2 GPT-5 53.8%
- 3 GPT-5 Chat 51.5%
- 4 GPT-4.1 47.4%
- 5 o3-mini 45.6%
- 6
Qwen3-32B 21.1%
- 7
DeepSeek-R1-Distill-Qwen-32B 10.5%
OdysseyBench-Neo overall
7 models we track have a score here.
- 1 o3 61.3%
- 2 GPT-5 Chat ↑1 57.6%
- 3 GPT-4.1 ↑1 56.6%
- 4 GPT-5 ↓2 56%
- 5 o3-mini 49.3%
- 6
Qwen3-32B 25.5%
- 7
DeepSeek-R1-Distill-Qwen-32B 19.2%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The OdysseyBench 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 OdysseyBench 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
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 15 months ago, in August 2025, for GPT-5. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (10 entries).
- Pass Models scored 7 models we track, plus 3 other entries on its leaderboard.
- Warning Newest score The newest came 15 months ago, in August 2025, for GPT-5.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5 (#71).
- Pass Headroom The best is 58.7%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 7 rows for models we track and 3 other entries 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.