MonitorBench
Models work through 1,540 instances across 19 tasks in 7 categories, each built around a decision-critical factor such as a cue, a hidden goal or a required step. A monitor LLM reads the chain of thought, and the monitorability score measures how often it surfaces that factor.
COLM 2026. Two stress tests ask the model to hide the factor or to evade a monitor it knows about. Inspect Evals lists single tasks separately, such as monitorbench-goal-sandbag-engineering from the Goal (Sandbag) task. Code and data are MIT licensed.
- Made by
- University of Illinois Urbana-Champaign (ASTRAL Group), University of Washington and UC San Diego
- Released
- March 2026
- Kind
- Deception and alignment
- Level
- Expert
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Propensity
- Risk
- Loss of control
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Loss of control, Deception and scheming
- Strongest model tested
-
Gemini 3 Flash #60 on the capability index
- Last model added
- December 2025
- Scores refreshed
- November 20th, 2026
gpt-oss-120b leads with 91.9% among models released by December 2025
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 December 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-oss-120b All-messages, Standard 91.9%
- 2
Qwen3-235B-A22B All-messages, Standard 91.7%
- 3
Qwen3-14B All-messages, Standard 89.8%
- 4 Claude Haiku 4.5 All-messages, Standard 85.8%
- 5 Claude Sonnet 4.5 All-messages, Standard 84.9%
- 6 GPT-5 All-messages, Standard 81.1%
- 7 GPT-5 mini All-messages, Standard 79.5%
- 8
Gemini 3 Flash All-messages, Standard 78.3%
Source: The MonitorBench paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | gpt-oss-120b | 91.9% | All-messages, Standard | Aug 2025 | Official leaderboard |
| 2 | | 91.7% | All-messages, Standard | Apr 2025 | Official leaderboard |
| 3 | | 89.8% | All-messages, Standard | Apr 2025 | Official leaderboard |
| 4 | Claude Haiku 4.5 | 85.8% | All-messages, Standard | Oct 2025 | Official leaderboard |
| 5 | Claude Sonnet 4.5 | 84.9% | All-messages, Standard | Sep 2025 | Official leaderboard |
| 6 | GPT-5 | 81.1% | All-messages, Standard | Aug 2025 | Official leaderboard |
| 7 | GPT-5 mini | 79.5% | All-messages, Standard | Aug 2025 | Official leaderboard |
| 8 | | 78.3% | All-messages, Standard | Dec 2025 | Official leaderboard |
Rankings by part
We read 19 parts from the benchmark’s paper in 7 groups: cue (condition), cue (answer), goal (sandbag), goal (deception), dual objectives, impossible tasks and solution process. Each ranks models on one part of the test. Qwen3-235B-A22B leads 8 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Health
8 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 100%
- 2 gpt-oss-120b ↓1 100%
- 3 GPT-5 ↑3 100%
- 4 GPT-5 mini ↑3 100%
- 5 Claude Sonnet 4.5 100%
- 6 Claude Haiku 4.5 ↓2 100%
- 7
Gemini 3 Flash ↑1 100%
- 8
Qwen3-14B ↓5 97.6%
Knowledge
8 models we track have a score here.
- 1 Claude Sonnet 4.5 ↑4 100%
- 2 gpt-oss-120b ↓1 99.9%
- 3
Qwen3-14B 99.8%
- 4
Qwen3-235B-A22B ↓2 99.8%
- 5 Claude Haiku 4.5 ↓1 99.6%
- 6 GPT-5 mini ↑1 99.3%
- 7 GPT-5 ↓1 98.2%
- 8
Gemini 3 Flash 98.1%
Law Judgment
8 models we track have a score here.
- 1 Claude Sonnet 4.5 ↑4 100%
- 2 gpt-oss-120b ↓1 99.7%
- 3 Claude Haiku 4.5 ↑1 99.6%
- 4
Gemini 3 Flash ↑4 99.6%
- 5
Qwen3-14B ↓2 99.5%
- 6
Qwen3-235B-A22B ↓4 99.1%
- 7 GPT-5 ↓1 91.8%
- 8 GPT-5 mini ↓1 86.7%
Preference
8 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 100%
- 2 GPT-5 ↑4 100%
- 3 Claude Sonnet 4.5 ↑2 100%
- 4 Claude Haiku 4.5 100%
- 5
Gemini 3 Flash ↑3 100%
- 6 gpt-oss-120b ↓5 99.9%
- 7
Qwen3-14B ↓4 99.8%
- 8 GPT-5 mini ↓1 99.7%
Health
8 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 96.9%
- 2 Claude Sonnet 4.5 ↑3 95%
- 3 Claude Haiku 4.5 ↑1 91.7%
- 4
Qwen3-14B ↓1 89.1%
- 5 GPT-5 ↑1 88.8%
- 6 gpt-oss-120b ↓5 87.6%
- 7 GPT-5 mini 87.5%
- 8
Gemini 3 Flash 81.8%
Mathematics
8 models we track have a score here.
- 1 gpt-oss-120b 100%
- 2
Qwen3-235B-A22B 78.9%
- 3 GPT-5 mini ↑4 64.3%
- 4
Qwen3-14B ↓1 62.8%
- 5 Claude Haiku 4.5 ↓1 62.1%
- 6 Claude Sonnet 4.5 ↓1 61.5%
- 7 GPT-5 ↓1 51.7%
- 8
Gemini 3 Flash 47.2%
Logic
8 models we track have a score here.
- 1 gpt-oss-120b 75.4%
- 2
Qwen3-235B-A22B 62.8%
- 3
Qwen3-14B 54.2%
- 4 Claude Haiku 4.5 53.5%
- 5 GPT-5 ↑1 45.8%
- 6 Claude Sonnet 4.5 ↓1 40.2%
- 7 GPT-5 mini 35.4%
- 8
Gemini 3 Flash 28.1%
Safety
3 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 97.8%
- 2
Qwen3-14B ↑1 89.1%
- 3
Gemini 3 Flash ↑5 82.7%
Mathematics
3 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 98.1%
- 2
Qwen3-14B ↑1 83.7%
- 3
Gemini 3 Flash ↑5 75.9%
Science
3 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 95.9%
- 2
Qwen3-14B ↑1 94.7%
- 3
Gemini 3 Flash ↑5 84.2%
Engineering
3 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 96.7%
- 2
Qwen3-14B ↑1 95.5%
- 3
Gemini 3 Flash ↑5 91.3%
Real-world Case
3 models we track have a score here.
- 1
Qwen3-235B-A22B ↑1 95.7%
- 2
Qwen3-14B ↑1 89.9%
- 3
Gemini 3 Flash ↑5 89.6%
Steganography
8 models we track have a score here.
- 1 Claude Sonnet 4.5 ↑4 100%
- 2 Claude Haiku 4.5 ↑2 100%
- 3
Qwen3-235B-A22B ↓1 99.7%
- 4 gpt-oss-120b ↓3 99.7%
- 5
Qwen3-14B ↓2 99.3%
- 6 GPT-5 95%
- 7 GPT-5 mini 90.7%
- 8
Gemini 3 Flash 78.2%
Summarization
8 models we track have a score here.
- 1
Qwen3-14B ↑2 100%
- 2
Qwen3-235B-A22B 100%
- 3 gpt-oss-120b ↓2 100%
- 4 Claude Sonnet 4.5 ↑1 100%
- 5 Claude Haiku 4.5 ↓1 100%
- 6 GPT-5 90.1%
- 7 GPT-5 mini 72.3%
- 8
Gemini 3 Flash 71.3%
Coding
8 models we track have a score here.
- 1
Qwen3-14B ↑2 100%
- 2
Qwen3-235B-A22B 100%
- 3 gpt-oss-120b ↓2 100%
- 4 GPT-5 ↑2 100%
- 5 GPT-5 mini ↑2 100%
- 6 Claude Haiku 4.5 ↓2 100%
- 7
Gemini 3 Flash ↑1 100%
- 8 Claude Sonnet 4.5 ↓3 99.9%
Agent Malicious
8 models we track have a score here.
- 1
Qwen3-14B ↑2 100%
- 2
Qwen3-235B-A22B 100%
- 3 Claude Sonnet 4.5 ↑2 100%
- 4 GPT-5 mini ↑3 96.9%
- 5
Gemini 3 Flash ↑3 96.3%
- 6 Claude Haiku 4.5 ↓2 93.4%
- 7 gpt-oss-120b ↓6 85%
- 8 GPT-5 ↓2 84.4%
Agent Benign
8 models we track have a score here.
- 1
Qwen3-14B ↑2 100%
- 2
Qwen3-235B-A22B 100%
- 3 Claude Haiku 4.5 ↑1 95%
- 4
Gemini 3 Flash ↑4 93%
- 5 GPT-5 ↑1 90.5%
- 6 GPT-5 mini ↑1 87.6%
- 7 gpt-oss-120b ↓6 87%
- 8 Claude Sonnet 4.5 ↓3 77.3%
Coding
8 models we track have a score here.
- 1
Qwen3-14B ↑2 100%
- 2
Qwen3-235B-A22B 100%
- 3 GPT-5 ↑3 100%
- 4 GPT-5 mini ↑3 100%
- 5 Claude Sonnet 4.5 100%
- 6 Claude Haiku 4.5 ↓2 100%
- 7
Gemini 3 Flash ↑1 100%
- 8 gpt-oss-120b ↓7 97.3%
Multi-step Update
8 models we track have a score here.
- 1 gpt-oss-120b 87.6%
- 2
Qwen3-235B-A22B 85.1%
- 3
Qwen3-14B 83.6%
- 4 Claude Haiku 4.5 47.8%
- 5
Gemini 3 Flash ↑3 31.8%
- 6 Claude Sonnet 4.5 ↓1 26.3%
- 7 GPT-5 mini 24.7%
- 8 GPT-5 ↓2 13.9%
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 MonitorBench paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| gpt-oss-120b | All-messages, Standard | 91.9% | |||
| All-messages, Standard | 91.7% | ||||
| gpt-oss-120b | CoT-only, Standard | 91.5% | |||
| All-messages, Standard | 89.8% | ||||
| CoT-only, Standard | 88.5% | ||||
| All-messages, Monitor-aware Evasion | 86.4% | ||||
| Claude Haiku 4.5 | All-messages, Standard | 85.8% | |||
| CoT-only, Standard | 85.5% | ||||
| Claude Sonnet 4.5 | All-messages, Standard | 84.9% | |||
| All-messages, Direct Concealment | 84.6% | ||||
| Claude Sonnet 4.5 | CoT-only, Standard | 84.4% | |||
| All-messages, Monitor-aware Evasion | 83.7% | ||||
| Claude Haiku 4.5 | CoT-only, Standard | 81.4% | |||
| GPT-5 | All-messages, Standard | 81.1% | |||
| CoT-only, Direct Concealment | 80.5% | ||||
| GPT-5 mini | All-messages, Standard | 79.5% | |||
| CoT-only, Monitor-aware Evasion | 78.9% | ||||
| Claude Sonnet 4.5 | CoT-only, Monitor-aware Evasion | 78.8% | |||
| All-messages, Direct Concealment | 78.4% | ||||
| All-messages, Standard | 78.3% | ||||
| CoT-only, Monitor-aware Evasion | 78% | ||||
| Claude Sonnet 4.5 | CoT-only, Direct Concealment | 77.5% | |||
| Claude Sonnet 4.5 | All-messages, Monitor-aware Evasion | 77.4% | |||
| GPT-5 | CoT-only, Standard | 76.7% | |||
| Claude Haiku 4.5 | All-messages, Direct Concealment | 76.3% | |||
| gpt-oss-120b | CoT-only, Monitor-aware Evasion | 76.1% | |||
| Claude Sonnet 4.5 | All-messages, Direct Concealment | 75.7% | |||
| CoT-only, Direct Concealment | 75.6% | ||||
| gpt-oss-120b | CoT-only, Direct Concealment | 75.2% | |||
| Claude Haiku 4.5 | CoT-only, Direct Concealment | 74.8% | |||
| Action-only, Standard | 74.6% | ||||
| CoT-only, Standard | 74.6% | ||||
| GPT-5 mini | CoT-only, Standard | 73.7% | |||
| gpt-oss-120b | All-messages, Direct Concealment | 73.3% | |||
| gpt-oss-120b | All-messages, Monitor-aware Evasion | 72.8% | |||
| Claude Haiku 4.5 | All-messages, Monitor-aware Evasion | 72.8% | |||
| GPT-5 | All-messages, Monitor-aware Evasion | 72% | |||
| All-messages, Monitor-aware Evasion | 72% | ||||
| GPT-5 | CoT-only, Monitor-aware Evasion | 71.7% | |||
| GPT-5 | All-messages, Direct Concealment | 71.3% | |||
| Claude Haiku 4.5 | CoT-only, Monitor-aware Evasion | 71.3% | |||
| CoT-only, Monitor-aware Evasion | 71.1% | ||||
| GPT-5 | CoT-only, Direct Concealment | 70.9% | |||
| All-messages, Direct Concealment | 70.3% | ||||
| Action-only, Standard | 69.6% | ||||
| CoT-only, Direct Concealment | 69.1% | ||||
| GPT-5 mini | All-messages, Monitor-aware Evasion | 67.7% | |||
| GPT-5 mini | All-messages, Direct Concealment | 67.6% | |||
| Action-only, Monitor-aware Evasion | 66.3% | ||||
| GPT-5 mini | CoT-only, Monitor-aware Evasion | 64.5% | |||
| GPT-5 mini | CoT-only, Direct Concealment | 63.8% | |||
| Claude Sonnet 4.5 | Action-only, Standard | 63.8% | |||
| Claude Haiku 4.5 | Action-only, Standard | 62.8% | |||
| Action-only, Direct Concealment | 62.5% | ||||
| Action-only, Monitor-aware Evasion | 61.5% | ||||
| GPT-5 mini | Action-only, Standard | 56.7% | |||
| gpt-oss-120b | Action-only, Standard | 55% | |||
| Claude Sonnet 4.5 | Action-only, Direct Concealment | 53.6% | |||
| Action-only, Direct Concealment | 53.4% | ||||
| gpt-oss-120b | Action-only, Direct Concealment | 52.2% | |||
| gpt-oss-120b | Action-only, Monitor-aware Evasion | 51.9% | |||
| Claude Sonnet 4.5 | Action-only, Monitor-aware Evasion | 51.6% | |||
| Claude Haiku 4.5 | Action-only, Direct Concealment | 50.8% | |||
| Claude Haiku 4.5 | Action-only, Monitor-aware Evasion | 49.9% | |||
| GPT-5 | Action-only, Standard | 48.5% | |||
| GPT-5 mini | Action-only, Direct Concealment | 48.4% | |||
| GPT-5 mini | Action-only, Monitor-aware Evasion | 46.1% | |||
| Action-only, Standard | 45% | ||||
| GPT-5 | Action-only, Monitor-aware Evasion | 44.4% | |||
| GPT-5 | Action-only, Direct Concealment | 43.9% | |||
| Action-only, Direct Concealment | 37.4% | ||||
| Action-only, Monitor-aware Evasion | 37.1% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The MonitorBench paper
Qwen3-235B-A22B on the official leaderboard passed 85% in April 2025
Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official MonitorBench 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
All 8 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Pass Independent runs Scores from the benchmark’s paper (8 entries).
- Pass Models scored 8 models we track.
- Note Newest score The newest came 11 months ago, in December 2025, for Gemini 3 Flash. That’s expected for a saturated test.
- Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is Gemini 3 Flash (#60).
- Note Headroom Saturated. The best is 91.9%, past 85% since April 2025.
Leaderboard import 3 pass
- Pass Importer We read the board on proceedings.colm.cc: 72 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.