Benchmark Almanac

OASIS (agent safety)

GPT-5 mini 95.2% leads since November 2025
Not in model cards

Agents handle multi-turn requests with different levels of complexity. Evaluators measure the proportion of safe responses.

Released
November 2025
Kind
Deception and alignment
Level
University
Scored
Settled by events
Answers
Agent task
Tasks
All public
Measures
Propensity
Behavior
Harmful use
Language
English
In and out
Text → Actions, Text
Interaction
Agentic
Items from
Generated
Refresh
One fixed set
Run by
Anyone
Strongest model tested
GPT-5 #71 on the capability index
Last model added
September 2025
Scores refreshed
November 20th, 2026
8
Entries in its paper, 7 models we track

GPT-5 mini leads with 95.2% among models released by September 2025 GPT-5 mini leads with 95.2% among models released by September 2025

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The headline is the refusal rate on harmful tasks. The false positive rate measures refusals on benign tasks. 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 September 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-5 mini Idealized Scenario, Greedy 95.2%
  2. 2 Qwen3-235B-A22B-Instruct Jul 2025, Idealized Scenario, Stochastic 94.1%
  3. 3 GPT-5 Idealized Scenario, Greedy 91.5%
  4. 4 Gemini 2.5 Pro (snapshot unspecified) Idealized Scenario, Stochastic 87.5%
  5. 5 Gemini 2.5 Flash Sep 2025, Idealized Scenario, Greedy 73.1%
  6. 6 Qwen3-235B-A22B-Thinking Jul 2025, Idealized Scenario, Stochastic 72.5%
  7. 7 DeepSeek-V3.1 Idealized Scenario, Greedy 63.2%
  8. 8 Qwen3-32B Idealized Scenario, Greedy 40%

Source: The OASIS (agent safety) paper

#EntryScoreSettingDateSource
1 GPT-5 mini 95.2% Idealized Scenario, Greedy Aug 2025 Official leaderboard
2 Qwen3-235B-A22B-Instruct Jul 2025 94.1% Idealized Scenario, Stochastic Jul 2025 Official leaderboard
3 GPT-5 91.5% Idealized Scenario, Greedy Aug 2025 Official leaderboard
4 Gemini 2.5 Pro (snapshot unspecified) 87.5% Idealized Scenario, Stochastic Official leaderboard
5 Gemini 2.5 Flash Sep 2025 73.1% Idealized Scenario, Greedy Sep 2025 Official leaderboard
6 Qwen3-235B-A22B-Thinking Jul 2025 72.5% Idealized Scenario, Stochastic Jul 2025 Official leaderboard
7 DeepSeek-V3.1 63.2% Idealized Scenario, Greedy Aug 2025 Official leaderboard
8 Qwen3-32B 40% Idealized Scenario, Greedy Apr 2025 Official leaderboard
#ModelScoreSettingDateSource
1 GPT-5 mini 95.2% Idealized Scenario, Greedy Aug 2025 Official leaderboard
2 Qwen3-235B-A22B-Instruct Jul 2025 94.1% Idealized Scenario, Stochastic Jul 2025 Official leaderboard
3 GPT-5 91.5% Idealized Scenario, Greedy Aug 2025 Official leaderboard
4 Gemini 2.5 Flash Sep 2025 73.1% Idealized Scenario, Greedy Sep 2025 Official leaderboard
5 Qwen3-235B-A22B-Thinking Jul 2025 72.5% Idealized Scenario, Stochastic Jul 2025 Official leaderboard
6 DeepSeek-V3.1 63.2% Idealized Scenario, Greedy Aug 2025 Official leaderboard
7 Qwen3-32B 40% Idealized Scenario, Greedy Apr 2025 Official leaderboard

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.

Lowest effort to highest, light to dark
GPT-5 mini
Realistic S… Idealized S…
95.2%
Qwen3-235B-A22B-Instruct
Realistic S… Idealized S…
94.1%
GPT-5
Realistic S… Idealized S…
91.5%
Gemini 2.5 Flash
Realistic S… Idealized S…
73.1%
Qwen3-235B-A22B-Thinking
Realistic S… Idealized S…
72.5%
DeepSeek-V3.1
Realistic S… Idealized S…
63.2%
Qwen3-32B
Realistic S… Idealized S…
40%

Source: The OASIS (agent safety) paper

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5 mini Idealized Scenario, Greedy 95.2%
Qwen3-235B-A22B-Instruct Jul 2025 Idealized Scenario, Stochastic 94.1%
Qwen3-235B-A22B-Instruct Jul 2025 Idealized Scenario, Greedy 93.9%
GPT-5 mini Idealized Scenario, Stochastic 93.3%
GPT-5 Idealized Scenario, Greedy 91.5%
GPT-5 Idealized Scenario, Stochastic 90.4%
GPT-5 Realistic Scenario, Greedy 77.9%
GPT-5 Realistic Scenario, Stochastic 76.3%
GPT-5 mini Realistic Scenario, Greedy 76%
GPT-5 mini Realistic Scenario, Stochastic 74.9%
Qwen3-235B-A22B-Instruct Jul 2025 Realistic Scenario, Greedy 74.4%
Qwen3-235B-A22B-Instruct Jul 2025 Realistic Scenario, Stochastic 73.6%
Gemini 2.5 Flash Sep 2025 Idealized Scenario, Greedy 73.1%
Qwen3-235B-A22B-Thinking Jul 2025 Idealized Scenario, Stochastic 72.5%
Gemini 2.5 Flash Sep 2025 Idealized Scenario, Stochastic 72%
Qwen3-235B-A22B-Thinking Jul 2025 Idealized Scenario, Greedy 70.1%
DeepSeek-V3.1 Idealized Scenario, Greedy 63.2%
DeepSeek-V3.1 Idealized Scenario, Stochastic 61.3%
Gemini 2.5 Flash Sep 2025 Realistic Scenario, Greedy 57.9%
Gemini 2.5 Flash Sep 2025 Realistic Scenario, Stochastic 57.1%
DeepSeek-V3.1 Realistic Scenario, Greedy 56.5%
DeepSeek-V3.1 Realistic Scenario, Stochastic 55.2%
Qwen3-32B Idealized Scenario, Greedy 40%
Qwen3-32B Idealized Scenario, Stochastic 38.9%
Qwen3-235B-A22B-Thinking Jul 2025 Realistic Scenario, Greedy 27.2%
Qwen3-235B-A22B-Thinking Jul 2025 Realistic Scenario, Stochastic 26.7%
Qwen3-32B Realistic Scenario, Greedy 16%
Qwen3-32B Realistic Scenario, Stochastic 15.7%

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.

Lowest effort to highest, light to dark
GPT-5 mini
Realistic S… Idealized S…
95.2%
Qwen3-235B-A22B-Instruct
Realistic S… Idealized S…
94.1%
GPT-5
Realistic S… Idealized S…
91.5%
Gemini 2.5 Flash
Realistic S… Idealized S…
73.1%
Qwen3-235B-A22B-Thinking
Realistic S… Idealized S…
72.5%
DeepSeek-V3.1
Realistic S… Idealized S…
63.2%
Qwen3-32B
Realistic S… Idealized S…
40%

Source: The OASIS (agent safety) paper

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5 mini Idealized Scenario, Greedy 95.2%
Qwen3-235B-A22B-Instruct Jul 2025 Idealized Scenario, Stochastic 94.1%
Qwen3-235B-A22B-Instruct Jul 2025 Idealized Scenario, Greedy 93.9%
GPT-5 mini Idealized Scenario, Stochastic 93.3%
GPT-5 Idealized Scenario, Greedy 91.5%
GPT-5 Idealized Scenario, Stochastic 90.4%
GPT-5 Realistic Scenario, Greedy 77.9%
GPT-5 Realistic Scenario, Stochastic 76.3%
GPT-5 mini Realistic Scenario, Greedy 76%
GPT-5 mini Realistic Scenario, Stochastic 74.9%
Qwen3-235B-A22B-Instruct Jul 2025 Realistic Scenario, Greedy 74.4%
Qwen3-235B-A22B-Instruct Jul 2025 Realistic Scenario, Stochastic 73.6%
Gemini 2.5 Flash Sep 2025 Idealized Scenario, Greedy 73.1%
Qwen3-235B-A22B-Thinking Jul 2025 Idealized Scenario, Stochastic 72.5%
Gemini 2.5 Flash Sep 2025 Idealized Scenario, Stochastic 72%
Qwen3-235B-A22B-Thinking Jul 2025 Idealized Scenario, Greedy 70.1%
DeepSeek-V3.1 Idealized Scenario, Greedy 63.2%
DeepSeek-V3.1 Idealized Scenario, Stochastic 61.3%
Gemini 2.5 Flash Sep 2025 Realistic Scenario, Greedy 57.9%
Gemini 2.5 Flash Sep 2025 Realistic Scenario, Stochastic 57.1%
DeepSeek-V3.1 Realistic Scenario, Greedy 56.5%
DeepSeek-V3.1 Realistic Scenario, Stochastic 55.2%
Qwen3-32B Idealized Scenario, Greedy 40%
Qwen3-32B Idealized Scenario, Stochastic 38.9%
Qwen3-235B-A22B-Thinking Jul 2025 Realistic Scenario, Greedy 27.2%
Qwen3-235B-A22B-Thinking Jul 2025 Realistic Scenario, Stochastic 26.7%
Qwen3-32B Realistic Scenario, Greedy 16%
Qwen3-32B Realistic Scenario, Stochastic 15.7%

Number one over time

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

OpenAI
Benchmark paper Passed 85% in Nov 2025, by GPT-5 mini
GPT-5 mini 95.2%

Source: The OASIS (agent safety) paper

Qwen3-235B-A22B-Instruct on the official leaderboard passed 85% in July 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.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) GPT-5 mini Qwen3-32B 0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) GPT-5 mini Qwen3-32B

Source: Official OASIS (agent safety) 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

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 7 models we track, plus 1 other entry on its leaderboard.
  • Note Newest score The newest came 14 months ago, in September 2025, for Gemini 2.5 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 GPT-5 (#71).
  • Note Headroom Saturated. The best is 95.2%, past 85% since July 2025.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 28 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.

How we check