LLM Position Bias Benchmark
Models judge pairs of responses twice, with the presentation order reversed.
The author publishes the model comparison in the repository.
- Made by
- Lech Mazur
- Kind
- Bias, fairness and politics
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Prediction
- Tasks
- All public
- Measures
- Propensity
- Behavior
- Unreliability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
- Claude Fable 5 #12 on the capability index
- Last model added
- June 2026
- Scores refreshed
- November 20th, 2026
MiMo-V2-Pro leads with 19.8% among models released by June 2026
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
The newest model with a score here came out in June 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 15 more
- 1 MiMo-V2-Pro 19.8%
- 2 Claude Fable 5 medium 27.3%
- 3 Claude Opus 4.8 high 27.4%
- 4
Seed2.0 Pro 28%
- 5
Gemini 3.5 Flash 29.8%
- 6 Claude Opus 4.6 high 30.2%
- 7
DeepSeek-V3.2 30.3%
- 8 GLM-5.1 31.5%
- 9
Qwen 3.6 Plus 33.8%
- 10
Qwen3.7-Max 34.8%
- 11
MiniMax-M3 34.9%
- 12
Gemini 3.1 Pro 35.4%
- 13
MiniMax-M2.7 36.5%
- 14
Trinity Large Thinking 36.6%
- 15 Claude Sonnet 4.6 high 37.4%
- 16 Claude Opus 4.7 high 37.7%
- 17 Grok 4.20 reasoning 39%
- 18
DeepSeek-V4-Pro 43.6%
- 19 GPT-5.5 medium 44.4%
- 20
ERNIE 5.0 45%
- 21
Gemini 3.1 Flash-Lite 45.3%
- 22 Kimi K2.6 47.3%
- 23
Qwen3.5 397B-A17B 48.5%
- 24
Gemma 4 31B IT 50.8%
- 25 Kimi K2.5 thinking 51.6%
- 26
Qwen3.5-122B-A10B 51.7%
- 27 GPT-5.4 Mini xhigh 54.7%
- 28 GPT-5.4 no reasoning 57.4%
- 29
Mistral Medium 3.1 57.9%
- 30
Mistral Large 3 58.7%
- 31
Mistral Medium 3.5 72.5%
Source: Official LLM Position Bias Benchmark leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | MiMo-V2-Pro | 19.8% | Mar 2026 | Official leaderboard | |
| 2 | Claude Fable 5 | 27.3% | medium | Jun 2026 | Official leaderboard |
| 3 | Claude Opus 4.8 | 27.4% | high | May 2026 | Official leaderboard |
| 4 | | 28% | Feb 2026 | Official leaderboard | |
| 5 | | 29.8% | May 2026 | Official leaderboard | |
| 6 | Claude Opus 4.6 | 30.2% | high | Feb 2026 | Official leaderboard |
| 7 | | 30.3% | Dec 2025 | Official leaderboard | |
| 8 | GLM-5.1 | 31.5% | Apr 2026 | Official leaderboard | |
| 9 | | 33.8% | Mar 2026 | Official leaderboard | |
| 10 | | 34.8% | May 2026 | Official leaderboard | |
| 11 | | 34.9% | Jun 2026 | Official leaderboard | |
| 12 | | 35.4% | Feb 2026 | Official leaderboard | |
| 13 | | 36.5% | Mar 2026 | Official leaderboard | |
| 14 | | 36.6% | Apr 2026 | Official leaderboard | |
| 15 | Claude Sonnet 4.6 | 37.4% | high | Feb 2026 | Official leaderboard |
| 16 | Claude Opus 4.7 | 37.7% | high | Apr 2026 | Official leaderboard |
| 17 | Grok 4.20 | 39% | reasoning | Feb 2026 | Official leaderboard |
| 18 | | 43.6% | Apr 2026 | Official leaderboard | |
| 19 | GPT-5.5 | 44.4% | medium | Apr 2026 | Official leaderboard |
| 20 | | 45% | Jan 2026 | Official leaderboard | |
| 21 | | 45.3% | Mar 2026 | Official leaderboard | |
| 22 | Kimi K2.6 | 47.3% | Apr 2026 | Official leaderboard | |
| 23 | | 48.5% | Feb 2026 | Official leaderboard | |
| 24 | | 50.8% | Apr 2026 | Official leaderboard | |
| 25 | Kimi K2.5 | 51.6% | thinking | Jan 2026 | Official leaderboard |
| 26 | | 51.7% | Feb 2026 | Official leaderboard | |
| 27 | GPT-5.4 Mini | 54.7% | xhigh | Mar 2026 | Official leaderboard |
| 28 | GPT-5.4 | 57.4% | no reasoning | Mar 2026 | Official leaderboard |
| 29 | | 57.9% | Aug 2025 | Official leaderboard | |
| 30 | | 58.7% | Dec 2025 | Official leaderboard | |
| 31 | | 72.5% | Apr 2026 | Official leaderboard |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: Official LLM Position Bias Benchmark leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| 72.5% | |||||
| GPT-5.4 | high | 66.3% | |||
| Grok 4.20 | non-reasoning | 64.8% | |||
| 58.7% | |||||
| 57.9% | |||||
| GPT-5.4 | no reasoning | 57.4% | |||
| reasoning | 54.9% | ||||
| GPT-5.4 Mini | xhigh | 54.7% | |||
| 51.7% | |||||
| Kimi K2.5 | thinking | 51.6% | |||
| 50.8% | |||||
| 48.5% | |||||
| Kimi K2.6 | 47.3% | ||||
| 45.3% | |||||
| 45% | |||||
| Claude Sonnet 4.6 | no reasoning | 44.8% | |||
| GPT-5.5 | medium | 44.4% | |||
| 43.6% | |||||
| Grok 4.20 | reasoning | 39% | |||
| Claude Opus 4.6 | no reasoning | 38.1% | |||
| Claude Opus 4.7 | high | 37.7% | |||
| Claude Sonnet 4.6 | high | 37.4% | |||
| 36.6% | |||||
| 36.5% | |||||
| 35.4% | |||||
| 34.9% | |||||
| 34.8% | |||||
| 33.8% | |||||
| GLM-5.1 | 31.5% | ||||
| 30.3% | |||||
| Claude Opus 4.6 | high | 30.2% | |||
| 29.8% | |||||
| 28% | |||||
| Claude Opus 4.8 | high | 27.4% | |||
| Claude Fable 5 | medium | 27.3% | |||
| MiMo-V2-Pro | 19.8% |
5 models have held number one on the official leaderboard since 2023
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official LLM Position Bias Benchmark leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official LLM Position Bias Benchmark leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
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%
- 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.
Gemini 4 Carbon 0%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
All 9 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (31 entries).
- Pass Models scored 31 models we track.
- Pass Newest score The newest is from June 2026, for Claude Fable 5.
- Note Frontier models None of the top 10 has a score. The strongest tested is Claude Fable 5 (#12).
- Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 36 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.