ReTrace
Computer agents recover from injected errors in desktop tasks. Replaying the remaining actions checks whether the recovered state produces the original successful outcome.
The framework uses OSWorld and WindowsAgentArena tasks. The repository still says code will be released soon, despite the paper describing availability.
- Made by
- KAIST, ETRI, and DeepAuto.ai
- Released
- October 2026
- Kind
- Desktop apps
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Agent task
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Desktop apps
- Items from
- Another benchmark
- Refresh
- One fixed set
- Run by
- A host
- Based on
- OSWorld, Windows Agent Arena
- In views
- Computer use, Agents
- Strongest model tested
- GPT-5.4 #28 on the capability index
- Last model added
- March 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
Recent advances in large (vision) language models have enabled computer-use agents that automate digital workflows via fine-grained keyboard and mouse operations. However, such agents remain fragile in long-horizon settings, where errors accumulate and executed actions leave unintended side effects, making the ability to recover from failures critical. Yet, existing remedies remain inadequate: environment resets are infeasible in proprietary settings, action replay is prohibitively expensive over long horizons, and API-based rollback is limited to web environments, leaving model-based recovery largely unexplored. To address this gap, we present ReTrace, a scalable evaluation framework for systematically studying recovery behavior in computer-use agents. We formalize Transition-Consistent Recovery: identical downstream actions from the recovered and target states must yield identical outcomes. We operationalize this criterion through goal-oriented action replay following controlled error injection. Using ReTrace, we present the first systematic study of autonomous, model-based recovery, revealing that (i) ideal recovery can unlock model's latent task completion capabilities, (ii) current computer-use agents fail to exploit their latent recovery capabilities, and (iii) accurate model-based recovery yields measurable end-to-end gains. Together, these findings establish recovery as a fundamental capability for building reliable computer-use agents. Our code and dataset is available at https://github.com/BangProx/ReTRACE.
GPT-5.4 leads with 63.6% among models released by March 2026 GPT-5.4 leads with 63.6% among models released by March 2026
Best score per entry 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. 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 March 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 GPT-5.4 Explicit, 3 steps with history through error 63.6%
- 2 Claude Sonnet 4.6 Explicit, 3 steps with history through error 62%
- 3
Qwen3.5-9B Explicit, 3 steps with recovery target contextualization 37.6%
- 4 EvoCUA-32B Explicit, 3 steps with recovery target contextualization 9.86%
Source: The ReTrace paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.4 | 63.6% | Explicit, 3 steps with history through error | Mar 2026 | Official leaderboard |
| 2 | Claude Sonnet 4.6 | 62% | Explicit, 3 steps with history through error | Feb 2026 | Official leaderboard |
| 3 | | 37.6% | Explicit, 3 steps with recovery target contextualization | Feb 2026 | Official leaderboard |
| 4 | EvoCUA-32B | 9.86% | Explicit, 3 steps with recovery target contextualization | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.4 | 63.6% | Explicit, 3 steps with history through error | Mar 2026 | Official leaderboard |
| 2 | Claude Sonnet 4.6 | 62% | Explicit, 3 steps with history through error | Feb 2026 | Official leaderboard |
| 3 | | 37.6% | Explicit, 3 steps with recovery target contextualization | Feb 2026 | Official leaderboard |
Rankings by part
We read 2 parts from the benchmark’s paper: OSWorld and WindowsAgentArena. Each ranks models on one part of the test. GPT-5.4 leads OSWorld and Claude Sonnet 4.6 leads WindowsAgentArena. Arrows show how far a model moved from its overall rank among the models we track.
OSWorld
3 models we track have a score here.
- 1 GPT-5.4 73.2%
- 2 Claude Sonnet 4.6 62.9%
- 3
Qwen3.5-9B 43.2%
WindowsAgentArena
3 models we track have a score here.
- 1 Claude Sonnet 4.6 ↑1 61.8%
- 2 GPT-5.4 ↓1 54.1%
- 3
Qwen3.5-9B 31.9%
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 ReTrace paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.4 | Explicit, 3 steps with history through error | 63.6% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps with history through error | 62% | |||
| GPT-5.4 | Explicit, 3 steps without history | 60.7% | |||
| GPT-5.4 | Explicit, 3 steps with recovery target contextualization | 57.3% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps with recovery target contextualization | 57.1% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps without history | 56.9% | |||
| GPT-5.4 | Explicit, 3 steps with history through recovery target | 55.8% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps with history through recovery target | 55.7% | |||
| Explicit, 3 steps with recovery target contextualization | 37.6% | ||||
| Claude Sonnet 4.6 | Implicit, 1 step | 31.9% | |||
| GPT-5.4 | Implicit, 1 step | 25.8% | |||
| Implicit, 1 step | 22.7% | ||||
| Claude Sonnet 4.6 | Implicit, 3 steps | 20.3% | |||
| Explicit, 3 steps with history through error | 15.3% | ||||
| Explicit, 3 steps without history | 11.8% | ||||
| Explicit, 3 steps with history through recovery target | 10.3% | ||||
| Implicit, 3 steps | 8.44% | ||||
| GPT-5.4 | Implicit, 3 steps | 6.27% |
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 ReTrace paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.4 | Explicit, 3 steps with history through error | 63.6% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps with history through error | 62% | |||
| GPT-5.4 | Explicit, 3 steps without history | 60.7% | |||
| GPT-5.4 | Explicit, 3 steps with recovery target contextualization | 57.3% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps with recovery target contextualization | 57.1% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps without history | 56.9% | |||
| GPT-5.4 | Explicit, 3 steps with history through recovery target | 55.8% | |||
| Claude Sonnet 4.6 | Explicit, 3 steps with history through recovery target | 55.7% | |||
| Explicit, 3 steps with recovery target contextualization | 37.6% | ||||
| Claude Sonnet 4.6 | Implicit, 1 step | 31.9% | |||
| GPT-5.4 | Implicit, 1 step | 25.8% | |||
| Implicit, 1 step | 22.7% | ||||
| Claude Sonnet 4.6 | Implicit, 3 steps | 20.3% | |||
| Explicit, 3 steps with history through error | 15.3% | ||||
| Explicit, 3 steps without history | 11.8% | ||||
| Explicit, 3 steps with history through recovery target | 10.3% | ||||
| Implicit, 3 steps | 8.44% | ||||
| GPT-5.4 | Implicit, 3 steps | 6.27% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The ReTrace 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 ReTrace 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 Computer and browser use
- BrowseComp In use Hard-to-find facts that need persistent web browsing. Short verifiable answers graded for exact correctness. Claude Fable 5 93.3%
- ScreenSpot-Pro In use GUI grounding on high-resolution professional-software screenshots. The model must locate the target UI element. GPT-6 Astra 92.7%
- AndroidWorld In use Agents complete 116 tasks across 20 real Android apps in an emulator. Each task checks the device's system state to judge success. GPT-6 Astra 100%
- OSWorld 2.1 In use Agents complete 108 long-horizon workflows across 31 self-hosted websites and desktop apps, averaging 318 tool calls. Scored by binary completion and checkpoint partial credit. Claude Opus 5 44.3%
- WebArena-Verified In use Agents complete 812 audited WebArena tasks on self-hosted websites. Deterministic evaluators check the agent's answer and captured network traces.
Muse Spark 1.1 69%
- OSWorld 2.1 (offline subset) In use Agents complete the 82 OSWorld 2.1 tasks on the official offline list, with no internet access. Scored by checkpoint partial credit and strict pass rate. Claude Fable 5.2 89.3%
Data audit
7 of 9 checks pass, 2 warnings. Look first at the Task access check. We don’t know whether the tasks are public. Every benchmark
Catalog record 1 warning 1 pass
- Warning Task access We don’t know whether the tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (4 entries).
- Note Models scored Only 3 models we track, plus 1 other entry on its leaderboard.
- Warning Newest score The newest came 8 months ago, in March 2026, for GPT-5.4.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.4 (#28).
- Pass Headroom The best is 63.6%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 18 rows for models we track and 1 other entry 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.