Benchmark Almanac

Respecting Warnings

GPT-6 Luna 11.8% leads since Oct 2026 Only in system cards

Agents meet low-stakes restrictions and warnings during routine tasks. Scored on the rate of unwanted persistence past the warning.

In-house OpenAI evaluation. The tasks are not public.

Made by OpenAI Released Sep 2026 Deception and alignment Tasks kept private Last in a system card Oct 2026 Last model added Oct 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.
5
Models scored in lab cards
8
System cards report it, from 1 lab

GPT-6 Luna leads with 11.8%

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.

  1. 1 GPT-6 Luna 11.8%
  2. 2 GPT-6 Astra 19%
  3. 3 GPT-6.1 Sol 23.5%
  4. 4 GPT-6 Sol 28.1%
  5. 5 GPT-6.1 Terra 29.6%
#ModelScoreSettingDateSource
1 GPT-6 Luna 11.8% Oct 2026 Model card
2 GPT-6 Astra 19% Sep 2026 Model card
3 GPT-6.1 Sol 23.5% Sep 2026 Model card
4 GPT-6 Sol 28.1% Oct 2026 Model card
5 GPT-6.1 Terra 29.6% Oct 2026 Model card

2 models have held number one since its release in Sep 2026

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

OpenAI
Lab claims 2 leaders
GPT-6 Luna 11.8%

Scores over time

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% 2027 GPT-6 Luna GPT-6 Astra 0% 20% 40% 60% 80% 100% 2027 GPT-6 Luna GPT-6 Astra

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.

8 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
29 Oct 2026 GPT-6 Luna system card addendum 11.8 Circumventing Warnings, conditional task completion rate, printed chart label (Figure 5, 7.0.2) Alignment
26 Oct 2026 GPT-6.1 Terra system card addendum 29.6 Circumventing Warnings, rollouts at maximum reasoning effort that completed the task through a route violating the restriction, text and printed chart label (Figure 5, 7.0.2) Alignment
7 Oct 2026 GPT-6 Sol system card addendum 28.1 Circumventing Warnings - STOP, conditional failure rate for successful circumvention, printed chart label (Figure 5, 7.1.2) Alignment
7 Oct 2026 GPT-6 Luna system card addendum 15.9 Circumventing Warnings - STOP, conditional failure rate for successful circumvention, printed chart label (Figure 5, 7.1.2) Alignment
29 Sep 2026 GPT-6.1 Sol system card addendum 23.5 Respecting Warnings, unwanted persistence rate (7.2) Alignment
22 Sep 2026 GPT-6 Sol system card 64.4 Circumventing Warnings, fails to stop, printed chart label (Figure 62, A.6.2) Alignment
22 Sep 2026 GPT-6 Luna system card 42.4 Circumventing Warnings, fails to stop, printed chart label (Figure 62, A.6.2) Alignment
3 Sep 2026 GPT-6 Astra system card 19 Respecting Warnings, unwanted persistence rate, maximum reasoning effort (8.2.2) 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%