Benchmark Almanac

Destructive action avoidance

GPT-5.5 90% leads since Apr 2026 Only in system cards

Coding agents meet conflicting user edits during tasks. Scored on preserving user changes and avoiding data-destructive actions.

In-house OpenAI evaluation for Codex models. The tasks are not public.

Made by OpenAI Released Feb 2026 Deception and alignment Tasks kept private Last in a system card Jul 2026 Last model added Jul 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Graded by a model A language model scores each response, usually against a rubric.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
8
Models scored in lab cards
11
System cards report it, from 1 lab

GPT-5.5 leads with 90% among models released by Jul 2026

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 Jul 2026, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 90.0 GPT-5.5
  2. 88.0 GPT-5.3 Codex
  3. 86.0 GPT-5.4
  4. 83.0 GPT-5.6 Sol
  5. 81.0 GPT-5.6 Terra
  6. 76.0 GPT-5.2 Codex
  7. 75.0 GPT-5.1-Codex-Max
  8. 73.0 GPT-5.6 Luna
  1. 1 GPT-5.5 90%
  2. 2 GPT-5.3 Codex 88%
  3. 3 GPT-5.4 86%
  4. 4 GPT-5.6 Sol 83%
  5. 5 GPT-5.6 Terra 81%
  6. 6 GPT-5.2 Codex 76%
  7. 7 GPT-5.1-Codex-Max 75%
  8. 8 GPT-5.6 Luna 73%
#ModelScoreSettingDateSource
1 GPT-5.5 90% Apr 2026 Model card
2 GPT-5.3 Codex 88% Feb 2026 Model card
3 GPT-5.4 86% Mar 2026 Model card
4 GPT-5.6 Sol 83% Jun 2026 Model card
5 GPT-5.6 Terra 81% Jun 2026 Model card
6 GPT-5.2 Codex 76% Dec 2025 Model card
7 GPT-5.1-Codex-Max 75% Nov 2025 Model card
8 GPT-5.6 Luna 73% Jun 2026 Model card

3 models have held number one since its release in Feb 2026

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

OpenAI
Lab claims Passed 85% in Feb 2026, by GPT-5.3 Codex
GPT-5.5 90%

The best score passed 85% in Feb 2026

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

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) GPT-5.5 GPT-5.3 Codex GPT-5.2 Codex 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) GPT-5.5 GPT-5.2 Codex

Labs report it only in system cards

We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.

11 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.

DateDocumentScoreAs the document writes itSection
9 Jul 2026 GPT-5.6 Terra system card 81 Avoiding Accidental Data-Destructive Actions, avoidance only 0.81 (avoidance + correctness 0.37) (3.3) Safety
9 Jul 2026 GPT-5.6 Sol system card 83 Avoiding Accidental Data-Destructive Actions, avoidance only 0.83 (avoidance + correctness 0.44) (3.3) Safety
9 Jul 2026 GPT-5.6 Luna system card 73 Avoiding Accidental Data-Destructive Actions, avoidance only 0.73 (avoidance + correctness 0.32) (3.3) Safety
25 Jun 2026 GPT-5.6 Terra system card 81 Avoiding Accidental Data-Destructive Actions, avoidance only 0.81 (avoidance + correctness 0.37) (3.3) Safety
25 Jun 2026 GPT-5.6 Sol system card 83 Avoiding Accidental Data-Destructive Actions, avoidance only 0.83 (avoidance + correctness 0.44) (3.3) Safety
25 Jun 2026 GPT-5.6 Luna system card 73 Avoiding Accidental Data-Destructive Actions, avoidance only 0.73 (avoidance + correctness 0.32) (3.3) Safety
23 Apr 2026 GPT-5.5 system card 90 Destructive action avoidance (Table 3, 3.3) Safety
5 Mar 2026 GPT-5.4 system card 86 Destructive action avoidance (Table 7, 3.7) Safety
5 Feb 2026 GPT-5.3 Codex system card 88 Destructive action avoidance, 0.88 (Table 2, 4.1.2) Alignment
18 Dec 2025 GPT-5.2 Codex system card addendum 76 Destructive action avoidance, 0.76 (Table 4, 4.2.2) Alignment
18 Nov 2025 GPT-5.1-Codex-Max system card 75 Destructive action avoidance, 0.75 (Table 6, 4.3.2.1) Alignment

More in Safety and alignment

All 444 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%
  • AgentDojo In use Agents do tool-use tasks while tools and data carry prompt injections. Attack success rate counts hijacked runs. Grok 4 Fast 0%
  • MASK In use Measures whether a model contradicts its own elicited beliefs under pressure to lie. Honesty separated from accuracy. Muse Spark Open 98.1%
  • 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. GPT-6 Astra 8.50%
  • Agentic misalignment suite (Anthropic) In use Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate. Claude Sonnet 4.5 0%