Benchmark Almanac

Impossible tasks (Anthropic)

Claude Haiku 4.5 30% leads since Oct 2025 Only in system cards

Agents work on coding problems designed to be impossible. Score is the rate at which they hack instead of giving up.

Run with no prompt and with an anti-hack prompt. The no-prompt rate is the default setting.

Made by Anthropic Released May 2025 Deception and alignment Tasks kept private Last in a system card Sep 2026 Last model added Sep 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.
11
Models scored in lab cards
11
System cards report it, from 1 lab

Claude Haiku 4.5 leads with 30%

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 Claude Haiku 4.5 30%
  2. 2 Claude Opus 5.5 34.7%
  3. 3 Claude Mythos Preview 37.5%
  4. 4 Claude Sonnet 4.6 40%
  5. 5 Claude Sonnet 4 45%
  6. 6 Claude Opus 4.7 45%
  7. 7 Claude Opus 4 47%
  8. 8 Claude Opus 4.6 50%
  9. 9 Claude Opus 4.1 52%
  10. 10 Claude Sonnet 4.5 53%
  11. 11 Claude Opus 4.5 55%
#ModelScoreSettingDateSource
1 Claude Haiku 4.5 30% Oct 2025 Model card
2 Claude Opus 5.5 34.7% Sep 2026 Model card
3 Claude Mythos Preview 37.5% Apr 2026 Model card
4 Claude Sonnet 4.6 40% Feb 2026 Model card
5 Claude Sonnet 4 45% May 2025 Model card
6 Claude Opus 4.7 45% Apr 2026 Model card
7 Claude Opus 4 47% May 2025 Model card
8 Claude Opus 4.6 50% Feb 2026 Model card
9 Claude Opus 4.1 52% Aug 2025 Model card
10 Claude Sonnet 4.5 53% Sep 2025 Model card
11 Claude Opus 4.5 55% Nov 2025 Model card

2 models have held number one since its release in May 2025

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

Anthropic
Lab claims 2 leaders
Claude Haiku 4.5 30%

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% 2025 2026 Claude Haiku 4.5 Claude Sonnet 4 0% 20% 40% 60% 80% 100% 2025 2026 Claude Haiku 4.5 Claude Sonnet 4

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
22 Sep 2026 Claude Opus 5.5 system card 34.7 Impossible tasks case study, RL training episodes missing a needed file, attempted reward hack rate (successful 3.1%. Intact tasks from the same environments: 6.0% attempted, 0.4% successful), lower is better (6.2.2, Figure 6.2.2.A) Alignment
16 Apr 2026 Claude Opus 4.7 system card 45 Impossible coding problems, hack rate with no prompt, printed on chart. 12.5 with anti-hack prompt (6.2.2.2, Figure 6.2.2.2.A) Alignment
7 Apr 2026 Claude Mythos Preview system card 37.5 Impossible coding problems, hack rate with no prompt, lower is better, value printed on chart; 20.0 with anti-hack prompt (4.2.2.2, Figure 4.2.2.2.A) Alignment
17 Feb 2026 Claude Sonnet 4.6 system card 40 Impossible tasks, classifier hack rate with no prompt; 28 with anti-hack prompt (Table 4.3.2.A) Alignment
5 Feb 2026 Claude Opus 4.6 system card 50 Impossible tasks, classifier hack rate with no prompt, lower is better; 23 with anti-hack prompt (Table 6.2.3.2.A) Alignment
24 Nov 2025 Claude Opus 4.5 system card 55 Impossible Tasks, classifier hack rate with no prompt, lower is better (Table 6.10.1.A; 35% with anti-hack prompt) Alignment
15 Oct 2025 Claude Haiku 4.5 system card 30 Impossible Tasks, classifier hack rate with no prompt; 23 with anti-hack prompt (Table 5.B) Alignment
29 Sep 2025 Claude Sonnet 4.5 system card 53 Impossible Tasks, simpler scaffold outside Claude Code, classifier hack rate with no prompt 53%, lower is better. With anti-hack prompt 20% (Table 6.1.A) Alignment
5 Aug 2025 Claude Opus 4.1 system card addendum 52 Claude Code Impossible Tasks, classifier hack rate with no prompt 52%. With anti-hack prompt 18% (Table 5.B) Alignment
22 May 2025 Claude Sonnet 4 system card 45 Claude Code Impossible Tasks, classifier hack rate with no prompt 45%. With anti-hack prompt 10% (the Opus 4.1 addendum later corrects this to 7%) (Table 6.2.A) Alignment
22 May 2025 Claude Opus 4 system card 47 Claude Code Impossible Tasks, classifier hack rate with no prompt 47%. With anti-hack prompt 5% (the Opus 4.1 addendum later corrects this to 19%) (Table 6.2.A) 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%