Self-preference bias (Anthropic)
Claude Opus 5.5 0.07 leads since Sep 2026 Only in system cards
Models grade transcripts attributed to themselves or to other models. Measures whether scores favor the model's own outputs.
In-house Anthropic evaluation. The tasks are not public.
Made by Anthropic Released Sep 2025 Bias, fairness and politics Tasks kept private Last in a system card Oct 2026
Last model added Oct 2026
- Level
- Everyday Most adults, without training
- Scored
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Open-ended Free-form text, documents, designs or media judged as a whole.
4
Models scored in lab cards
10
System cards report it, from 1 lab
Claude Opus 5.5 leads with 0.07
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 Claude Opus 5.5 0.07
- 2 Claude Sonnet 5.5 0.16
- 3 Claude Haiku 5.5 0.23
- 4 Claude Opus 4.7 0.40
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 0.07 | Sep 2026 | Model card | |
| 2 | Claude Sonnet 5.5 | 0.16 | Sep 2026 | Model card | |
| 3 | Claude Haiku 5.5 | 0.23 | Oct 2026 | Model card | |
| 4 | Claude Opus 4.7 | 0.40 | Apr 2026 | Model card |
2 models have held number one since its release in Sep 2025
Lab claims count from the day the card came out, independent runs from the model’s release.
Anthropic
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
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.
10 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 |
|---|---|---|---|---|
| 7 Oct 2026 | Claude Haiku 5.5 system card | 0.23 | Self-preference bias, 0.23 points on a 0 to 9 scale with the Claude system prompt reminder, 0.15 without (6.3.2) | Alignment |
| 28 Sep 2026 | Claude Sonnet 5.5 system card | 0.16 | Self-preference bias, 0.16 points on a 0 to 9 scale when reminded it is Claude (0.06 without the reminder) (6.3.1, Figure 6.3.1.A) | Alignment |
| 22 Sep 2026 | Claude Opus 5.5 system card | 0.07 | Self-preference bias, 0.07 points on a 10-point scale with the Claude-identity system prompt (0.01 with no system prompt) (6.5.3, Figure 6.5.3.A) | Alignment |
| 28 May 2026 | Claude Opus 4.8 system card | Self-preference bias in four prompting conditions on a 10-point scale, chart without printed values. Text says zero is within the error bars in all four (6.3.5, Figure 6.3.5.A) | Alignment | |
| 16 Apr 2026 | Claude Opus 4.7 system card | 0.4 | Self-preference bias, 0.4 on a 10-point scale in the largest condition (Claude system prompt, CoT on), stated in text. Other three conditions chart without printed values (6.3.5, Figure 6.3.5.A) | Alignment |
| 7 Apr 2026 | Claude Mythos Preview system card | Self-recognition bias in transcript grading, four conditions, chart without printed values. Text says an earlier version showed less than 0.1 points on a 10-point scale (4.3.5, Figure 4.3.5.A) | Alignment | |
| 17 Feb 2026 | Claude Sonnet 4.6 system card | Self-preference bias in transcript grading, four conditions, chart without printed values (Figure 4.6.2.A) | Alignment | |
| 5 Feb 2026 | Claude Opus 4.6 system card | Self-preference evaluation, score difference when transcripts are labeled as written by Claude, four conditions, chart without printed values; text says less than 0.2 points on a 10-point scale (Figure 6.3.7.A) | Alignment | |
| 24 Nov 2025 | Claude Opus 4.5 system card | Self-preference evaluation, Benchmark and Poems variants, change in probability of choosing itself, chart without printed values (Figure 6.8.A) | Alignment | |
| 29 Sep 2025 | Claude Sonnet 4.5 system card | Self-preference on decision tasks, Benchmark and Poems variants, increase in probability of choosing itself, chart without printed values (Figure 7.5.9.1.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%