Benchmark Almanac

ASPI

Gemini 3.1 Pro 17.8% leads since May 2026
Not in model cards

Agents complete simulated tasks with incomplete instructions. The authors compare adherence to user intent across matched versions of each task.

The paper reports results for ten models on 728 scenarios. Its published measure is attack success, so lower is better.

Made by
Scale AI
Released
May 2026
Kind
Deception and alignment
Level
Everyday
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Propensity
Language
English
In and out
Text → Actions, Text
Interaction
Agentic
Items from
Written by experts
Refresh
One fixed set
Run by
Anyone
In views
Agents, Robotics
Strongest model tested
GPT-5.5 #21 on the capability index
Last model added
April 2026
Scores refreshed
November 20th, 2026
10
Models scored in its paper

Gemini 3.1 Pro leads with 17.8% among models released by April 2026

Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The headline averages attack success across six conditions, giving each condition equal weight. Where a model ran with several settings or harnesses, we keep its best.

The newest model with a score here came out in April 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 Gemini 3.1 Pro Prompt guard 17.8%
  2. 2 Gemini 3 Flash Prompt guard 21.3%
  3. 3 o3 No defense 25.6%
  4. 4 GPT-5.5 No defense 26.9%
  5. 5 GPT-5.4 No defense 31.7%
  6. 6 Qwen3-32B No defense 33.7%
  7. 7 Claude Opus 4.7 No defense 34.3%
  8. 8 Kimi K2.5 No defense 51.9%
  9. 9 Qwen3-235B-A22B-Instruct Jul 2025, No defense 53.3%
  10. 10 DeepSeek-V3.2 No defense 68.3%

Source: The ASPI paper

#ModelScoreSettingDateSource
1 Gemini 3.1 Pro 17.8% Prompt guard Feb 2026 Official leaderboard
2 Gemini 3 Flash 21.3% Prompt guard Dec 2025 Official leaderboard
3 o3 25.6% No defense Apr 2025 Official leaderboard
4 GPT-5.5 26.9% No defense Apr 2026 Official leaderboard
5 GPT-5.4 31.7% No defense Mar 2026 Official leaderboard
6 Qwen3-32B 33.7% No defense Apr 2025 Official leaderboard
7 Claude Opus 4.7 34.3% No defense Apr 2026 Official leaderboard
8 Kimi K2.5 51.9% No defense Jan 2026 Official leaderboard
9 Qwen3-235B-A22B-Instruct Jul 2025 53.3% No defense Jul 2025 Official leaderboard
10 DeepSeek-V3.2 68.3% No defense Dec 2025 Official leaderboard

Rankings by part

We read 30 parts from the benchmark’s paper, each ranking models on one part of the test, such as exec_tool, exec_user and exec_next. Claude Opus 4.7 leads 12 of them, and 5 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

exec_tool

10 models we track have a score here.

  1. 1 Gemini 3.1 Pro 0%
  2. 2 GPT-5.4 ↑3 0%
  3. 3 GPT-5.5 ↑1 0%
  4. 4 Gemini 3 Flash ↓2 0.30%
  5. 5 o3 ↓2 1.80%
  6. 6 Claude Opus 4.7 ↑1 2.10%
  7. 7 Kimi K2.5 ↑1 11.1%
  8. 8 Qwen3-32B ↓2 18.7%
  9. 9 Qwen3-235B-A22B-Instruct Jul 2025 41.9%
  10. 10 DeepSeek-V3.2 65.4%

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Google DeepMind
Benchmark paper Not tested since May 2026

Source: The ASPI paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Gemini 3.1 Pro o3 0% 20% 40% 60% 80% 100% 2025 2026 Gemini 3.1 Pro o3

Source: Official ASPI 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

All 636 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 7 months ago, in April 2026, for GPT-5.5. 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 (10 entries).
  • Pass Models scored 10 models we track.
  • Warning Newest score The newest came 7 months ago, in April 2026, for GPT-5.5.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
  • 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: 14 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.

How we check