Agentic tasks (autonomous replication and adaptation)
Agents attempt resource acquisition tasks such as building an authenticated API proxy or loading a model on a cloud account. Scored as success rate.
In-house OpenAI Preparedness evaluation. Includes simplified variants with hints.
- Made by
- OpenAI
- Released
- August 2024
- Kind
- Labs' internal R&D tests
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Kept private
- Strongest model tested
- o3-mini #124 on the capability index
- Last in a system card
- February 2025
- Last model added
- February 2025
OpenAI deep research leads with 78% among models released by February 2025
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
The newest model with a score here came out in February 2025, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- Claude Fable 5.2
-
Gemini 4 Carbon - GPT-6 Astra
- Claude Opus 5.5
- and 16 more
- 1 OpenAI deep research 78%
- 2 GPT-4.5 40%
- 3 o3-mini 27%
- 4 GPT-4o May 2024 0%
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | OpenAI deep research | 78% | Feb 2025 | Lab model card | |
| 2 | GPT-4.5 | 40% | Feb 2025 | Lab model card | |
| 3 | o3-mini | 27% | Jan 2025 | Lab model card | |
| 4 | GPT-4o May 2024 | 0% | Aug 2024 | Lab model card |
3 models have held number one in labs’ own numbers since its release in August 2024
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.
8 system cards report it, from 1 lab
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Feb 27, 2025 | GPT-4.5 system card | 40 | Agentic Tasks, post-mitigation (4.6.3) | Autonomy |
| Feb 25, 2025 | OpenAI deep research system card | 78 | Agentic Tasks, agentic subtasks pass@1, post-mitigation with browsing, printed chart label (without browsing 80%) (3.4.7) | Autonomy |
| Jan 31, 2025 | o3-mini system card | 27 | Agentic Tasks, pass@1, o3-mini (Post), printed chart label (5.7.3) | Autonomy |
| Jan 23, 2025 | OpenAI Computer-Using Agent (CUA) system card | Autonomous replication eval in computer use, pre-mitigation Operator model, main tasks OpenAI API proxy 0%, GPU acquisition 60%, Mistral 7B docker 0% (prompt engineered 0%, 80%, 0%), no overall score (Table 2, 3.3.2) | Autonomy | |
| Dec 5, 2024 | o1 system card | Agentic Tasks, success rate over 100 trials on 10 tasks plus buy_gpu with Ranger browsing (8%), no overall score (5.8.3 charts) | Autonomy | |
| Sep 12, 2024 | o1-preview system card | Agentic Tasks: Success rates over 100 trials on 11 tasks, chart without printed numbers (4.5.3) | Autonomy | |
| Sep 12, 2024 | o1-mini system card | Agentic Tasks: Success rates over 100 trials on 11 tasks, chart without printed numbers (4.5.3) | Autonomy | |
| Aug 8, 2024 | GPT-4o system card | 0 | Autonomous replication and adaptation (ARA) tasks, 100 trials (3.8) | Autonomy |
More in AI research
- PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics.
Qwen 3.8 Max 93% - MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
- PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
- MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6 Astra 93.8%
- KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup.
MiniMax-M3 28.8% - PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours.
Inkling 2 Small 494.3