Average in-eval reward hacking rate (Anthropic)
Anthropic averages hack rates across its reward hacking evaluations into one number per model.
Summary of the reward-hack-prone coding, impossible tasks and training distribution results.
- 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.
Claude Sonnet 4 leads with 14.3% among models released by Oct 2025
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 Oct 2025, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6 Astra
- Claude Opus 5.5
- GPT-6.1 Terra
- GPT-6.1 Sol
- and 16 more
- 1 Claude Sonnet 4 14.3%
- 2 Claude Opus 4 15.2%
- 3 Claude Sonnet 4.5 15.4%
- 4 Claude Haiku 4.5 15.5%
- 5 Claude Opus 4.1 18.2%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 4 | 14.3% | May 2025 | Model card | |
| 2 | Claude Opus 4 | 15.2% | May 2025 | Model card | |
| 3 | Claude Sonnet 4.5 | 15.4% | Sep 2025 | Model card | |
| 4 | Claude Haiku 4.5 | 15.5% | Oct 2025 | Model card | |
| 5 | Claude Opus 4.1 | 18.2% | Aug 2025 | Model card |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
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.
5 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.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 15 Oct 2025 | Claude Haiku 4.5 system card | 15.5 | Average In-Eval Hack Rate across reward hacking evaluations, value printed on chart (Figure 5.A) | Alignment |
| 29 Sep 2025 | Claude Sonnet 4.5 system card | 15.4 | Average In-Eval Hack Rates 15.4%, lower is better (Figure 6.A) | Alignment |
| 5 Aug 2025 | Claude Opus 4.1 system card addendum | 18.2 | Averaged reward hacking rate 18.2%, lower is better (Figure 5.A) | Alignment |
| 22 May 2025 | Claude Sonnet 4 system card | 14.3 | Average In-Eval Hack Rates 14.3%, lower is better (Figure 6.A) | Alignment |
| 22 May 2025 | Claude Opus 4 system card | 15.2 | Average In-Eval Hack Rates 15.2%, lower is better (Figure 6.A) | Alignment |
More 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%