BBQ (Bias Benchmark for QA) — HELM safety-suite component; Ambig/Disambig Accuracy/Bias sub-metrics kept separate (metric axis)
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | Claude 4 Opus | Anthropic | 99.3 | 2 | 2026-06-18 |
| 2 | GPT OSS 120B | OpenAI | 98.5 | 1 | 2026-05-10 |
| 3 | o3 | OpenAI | 97.9 | 1 | 2026-05-10 |
| 4 | Claude Sonnet 4 | Anthropic | 97.9 | 3 | 2026-06-21 |
| 5 | GLM-4.5-Air | Z.ai | 97.8 | 1 | 2026-05-10 |
| 6 | Gemini 2.5 Flash | 97.7 | 1 | 2026-05-10 | |
| 7 | o1 | OpenAI | 97.3 | 2 | 2026-06-18 |
| 8 | OpenAI o1-mini | OpenAI | 96.9 | 1 | 2026-05-10 |
| 9 | Grok 3 Mini Beta | SpaceXAI | 96.7 | 1 | 2026-05-10 |
| 10 | DeepSeek-V3 | DeepSeek | 96.7 | 1 | 2026-05-10 |
| 11 | DeepSeek-R1 | DeepSeek | 96.6 | 2 | 2026-05-10 |
| 12 | Qwen3 235B A22B FP8 Throughput | Alibaba | 96.6 | 1 | 2026-05-10 |
| 13 | Gemini 2.5 Pro | 96.4 | 1 | 2026-05-10 | |
| 14 | Gemini 2.0 Pro (02-05 preview) | 95.6 | 1 | 2026-05-10 | |
| 15 | Palmyra-X-004 | Writer | 95.5 | 1 | 2026-05-10 |
| 16 | Qwen2.5 Instruct 72B | Alibaba | 95.4 | 1 | 2026-05-10 |
| 17 | Gemini 2.0 Flash (Reasoning) | 95.4 | 1 | 2026-06-24 | |
| 18 | Llama 3.1 Instruct Turbo (70B) | Meta | 95.4 | 1 | 2026-05-10 |
| 19 | Qwen2 Instruct (72B) | Alibaba | 95.1 | 1 | 2026-05-10 |
| 20 | GPT-4o | OpenAI | 95.1 | 1 | 2026-05-10 |
| 21 | Gemini 2.5 Flash Lite | 94.9 | 1 | 2026-05-10 | |
| 22 | Claude 3.5 Sonnet | Anthropic | 94.9 | 1 | 2026-05-10 |
| 23 | Kimi K2 Instruct | Moonshot | 94.9 | 1 | 2026-05-10 |
| 24 | Palmyra X5 | Writer | 94.8 | 1 | 2026-05-10 |
| 25 | Gemini 1.5 Flash (001) | 94.7 | 1 | 2026-05-10 | |
| 26 | Gemini 1.5 Pro | 94.5 | 1 | 2026-05-10 | |
| 27 | Llama 3.1 Instruct 405B | Meta | 94.5 | 1 | 2026-05-10 |
| 28 | Palmyra Fin | Writer | 94.2 | 1 | 2026-05-10 |
| 29 | GPT-4 Turbo | OpenAI | 94.1 | 1 | 2026-05-10 |
| 30 | o3-mini | OpenAI | 94.1 | 1 | 2026-05-10 |
| 31 | Claude 3 Opus | Anthropic | 94 | 1 | 2026-05-10 |
| 32 | o4-mini | OpenAI | 94 | 1 | 2026-05-10 |
| 33 | Grok 4 | SpaceXAI | 93.7 | 1 | 2026-05-10 |
| 34 | Grok 3 Beta | SpaceXAI | 93.6 | 1 | 2026-05-10 |
| 35 | Mistral Small 3 (Non-Reasoning) | Mistral | 93.3 | 2 | 2026-06-19 |
| 36 | Llama 4 Maverick 17B 128E Instruct FP8 | Meta | 93 | 1 | 2026-05-10 |
| 37 | GPT-4.1 | OpenAI | 92.6 | 1 | 2026-05-10 |
| 38 | GPT-4.1 mini | OpenAI | 92.1 | 2 | 2026-06-19 |
| 39 | Claude 3.7 Sonnet | Anthropic | 92.1 | 2 | 2026-06-21 |
| 40 | GPT-4.5 Preview | OpenAI | 92 | 1 | 2026-05-10 |
| 41 | Gemini 2.0 Flash Lite (02-05 preview) | 92 | 1 | 2026-06-24 | |
| 42 | Gemini 2.0 Flash (Non-Reasoning) | 92 | 1 | 2026-05-10 | |
| 43 | Llama 3 Instruct 70B | Meta | 91 | 2 | 2026-06-19 |
| 44 | Qwen2.5 Instruct Turbo (7B) | Alibaba | 90.6 | 1 | 2026-05-10 |
| 45 | Claude 3 Sonnet | Anthropic | 90 | 1 | 2026-05-10 |
| 46 | Command R+ (Apr '24) | Cohere | 89.9 | 1 | 2026-05-10 |
| 47 | GPT-4o mini | OpenAI | 88.2 | 1 | 2026-05-10 |
| 48 | Llama 4 Scout (17Bx16E) Instruct | Meta | 87.5 | 1 | 2026-06-26 |
| 49 | GPT-4.1 nano | OpenAI | 87.5 | 2 | 2026-06-19 |
| 50 | Llama 4 Scout | Meta | 87.5 | 1 | 2026-05-10 |
| 51 | DeepSeek-Chat | DeepSeek | 86.2 | 1 | 2026-05-10 |
| 52 | Granite 3.3 8B Instruct | IBM | 84 | 1 | 2026-05-10 |
| 53 | Palmyra Med | Writer | 80.1 | 1 | 2026-05-10 |
| 54 | DBRX Instruct | Databricks | 79.2 | 1 | 2026-05-10 |
| 55 | Llama 3.1 Instruct Turbo (8B) | Meta | 78.5 | 1 | 2026-05-10 |
| 56 | Llama 3 Instruct 8B | Meta | 76.5 | 2 | 2026-06-19 |
| 57 | Command R | Cohere | 72.4 | 1 | 2026-05-10 |
| 58 | OLMo 2 32B Instruct March 2025 | AllenAI | 71.4 | 1 | 2026-05-10 |
| 59 | OLMo 2 13B Instruct November 2024 | AllenAI | 70.4 | 1 | 2026-05-10 |
| 60 | Mistral Instruct V0.3 (7B) | Mistral | 66.8 | 1 | 2026-05-10 |
| 61 | Claude 3 Haiku | Anthropic | 62.5 | 1 | 2026-05-10 |
| 62 | Mistral V0.1 (7B) | Mistral | 56.5 | 1 | 2026-05-10 |
| 63 | OLMo 2 7B Instruct November 2024 | AllenAI | 52.5 | 1 | 2026-05-10 |
| 64 | OLMoE 1B-7B Instruct January 2025 | AllenAI | 41.9 | 1 | 2026-05-10 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.