InstrumentalEval
Models respond to scenarios where gaining resources or preserving themselves could serve another goal. A grader identifies responses that pursue those intermediate goals.
- Made by
- Yufei He and colleagues
- Released
- February 2025
- Kind
- Deception and alignment
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Propensity
- Risk
- Loss of control
- Behavior
- Power-seeking
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Loss of control, Deception and scheming
- Strongest model tested
- o1 #118 on the capability index
- Last model added
- January 2025
- Scores refreshed
- November 20th, 2026
GPT-4o (snapshot unspecified) tops the leaderboard with 15.8% o1-mini leads with 43.4% among models released by January 2025
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. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is o1-mini, at 43.4%. This view shows only models we track, each at its best.
The newest model with a score here came out in January 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 GPT-4o (snapshot unspecified) Without goal nudging 15.8%
- 2 DeepSeek-V3 (snapshot unspecified) With goal nudging 21.1%
- 3 Sonnet-3.5 With goal nudging 22.4%
- 4 Gemini 2.0 FTE. With goal nudging 32.9%
- 5 DeepSeek-R1 (snapshot unspecified) Without goal nudging 34.2%
- 6 o1-mini With goal nudging 43.4%
- 7 o1 With goal nudging 44.7%
- 8 o3-mini With goal nudging 48.7%
Source: The InstrumentalEval paper
- 1 o1-mini With goal nudging 43.4%
- 2 o1 With goal nudging 44.7%
- 3 o3-mini With goal nudging 48.7%
Source: The InstrumentalEval paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-4o (snapshot unspecified) | 15.8% | Without goal nudging | Official leaderboard | |
| 2 | DeepSeek-V3 (snapshot unspecified) | 21.1% | With goal nudging | Official leaderboard | |
| 3 | Sonnet-3.5 | 22.4% | With goal nudging | Official leaderboard | |
| 4 | Gemini 2.0 FTE. | 32.9% | With goal nudging | Official leaderboard | |
| 5 | DeepSeek-R1 (snapshot unspecified) | 34.2% | Without goal nudging | Official leaderboard | |
| 6 | o1-mini | 43.4% | With goal nudging | Sep 2024 | Official leaderboard |
| 7 | o1 | 44.7% | With goal nudging | Dec 2024 | Official leaderboard |
| 8 | o3-mini | 48.7% | With goal nudging | Jan 2025 | 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 Evading Shutdown, Hacking Computer Systems and Running Multiple AI Copies. o1-mini leads 5 of them, and o1 leads the other one. Arrows show how far a model moved from its overall rank among the models we track.
Running Multiple AI Copies
3 models we track have a score here.
Hiding Unwanted Behavior
3 models we track have a score here.
Strategically Appearing Aligned
3 models we track have a score here.
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The InstrumentalEval 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 InstrumentalEval 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 Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6%
- Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts.
Gemini 4 Carbon 0%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 22 months ago, in January 2025, for o3-mini. Every benchmark
Catalog record 2 pass
- Pass Task access All 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 (8 entries).
- Pass Models scored 3 models we track, plus 5 other entries on its leaderboard.
- Warning Newest score The newest came 22 months ago, in January 2025, for o3-mini.
- Note Frontier models None of the top 10 has a score. The strongest tested is o1 (#118).
- Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 3 rows for models we track and 5 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.