HELM Capabilities mean
Mean of HELM's five capability scenarios. Shown for reference.
As of 12 Sept 2026, GPT-5 mini leads HELM Capabilities mean on BenchLeader with 81.9%, ahead of o4-mini at 81.2%, across 68 model configurations with a published result.
- Published by
- HELM Capabilities
- Category
- Composite
- Index weight
- Reference only
- Models
- 68
- Data as of
- 12 Sept 2026
Stanford Center for Research on Foundation Models (crfm.stanford.edu), Apache-2.0 results.
What the test looks like
The mean of HELM's five capability scenarios: MMLU-Pro, GPQA, IFEval, WildBench and Omni-MATH.
How it is scored
Unweighted mean of the five scores, on a 0 to 100 scale. Shown for reference only.
What to keep in mind
Five scenarios is a narrow base for a composite, and HELM updates roughly monthly, so new models appear with a lag.
- 1GPT-5 mini81.9%
- 2o4-mini81.2%
- 3o381.1%
- 4GPT-580.7%
- 5Gemini 3 Pro79.9%
- 6Qwen3 235B A22B 250779.8%
- 7Grok 478.5%
- 8Claude Opus 478.0%
- 9gpt-oss-120b77.0%
- 10Kimi K276.8%
- 11Claude Sonnet 476.6%
- 12Claude Sonnet 4.576.2%
- 13Claude Opus 475.7%
- 14GPT-5 nano74.8%
- 15Gemini 2.5 Pro74.5%
68 of 68
| # | |||||
|---|---|---|---|---|---|
| 1 |  GPT-5 miniOpenAI | 81.9% | – | 57.9 | – |
| 2 | o4-miniOpenAI | 81.2% | – | 57.2 | – |
| 3 | o3OpenAI | 81.1% | – | 61.6 | – |
| 4 | GPT-5OpenAI | 80.7% | – | 60.8 | – |
| 5 |  Gemini 3 ProGoogle | 79.9% | – | 62.0 | – |
| 6 |  Qwen3 235B A22B 2507Alibabaopen | 79.8% | Qwen3 235B A22B Instruct 2507 FP8 | 54.1 | – |
| 7 |  Grok 4xAI | 78.5% | 0709 | 58.6 | – |
| 8 |  Claude Opus 4Anthropic | 78.0% | – | 56.1 | – |
| 9 | gpt-oss-120bOpenAIopen | 77.0% | – | 48.8 | – |
| 10 |  Kimi K2Moonshot AIopen | 76.8% | Kimi K2 Instruct | 51.2 | – |
| 11 | Claude Sonnet 4Anthropic | 76.6% | – | 55.1 | – |
| 12 | Claude Sonnet 4.5Anthropic | 76.2% | 20250929 | 53.7 | – |
| 13 | Claude Opus 4Anthropic | 75.7% | 20250514 | 52.2 | – |
| 14 | GPT-5 nanoOpenAI | 74.8% | – | 52.5 | – |
| 15 | Gemini 2.5 ProGoogle | 74.5% | – | 54.5 | – |
| 16 | Claude Sonnet 4Anthropic | 73.3% | 20250514 | 50.5 | – |
| 17 | Grok 3xAI | 72.7% | – | 51.0 | – |
| 18 | GPT-4.1OpenAI | 72.7% | – | 48.8 | – |
| 19 | GPT-4.1 miniOpenAI | 72.6% | – | 46.3 | – |
| 20 | Qwen3 235B A22BAlibabaopen | 72.6% | Qwen3 235B A22B FP8 Throughput | 44.8 | – |
| 21 |  Llama 4 MaverickMetaopen | 71.8% | Llama 4 Maverick (17Bx128E) Instruct FP8 | 44.2 | – |
| 22 | Claude Haiku 4.5Anthropic | 71.7% | 20251001 | 48.6 | – |
| 23 | Qwen3-Next 80B A3BAlibabaopen | 70.0% | – | 51.2 | – |
| 24 |  DeepSeek R1 0528DeepSeekopen | 69.9% | – | 49.9 | – |
| 25 |  Palmyra X5Writer | 69.6% | Palmyra X5 | – | – |
| 26 | Grok 3 minixAI | 67.9% | – | 51.8 | – |
| 27 | Gemini 2.0 FlashGoogle | 67.9% | Gemini 2.0 Flash | 43.5 | – |
| 28 | Claude 3.7 SonnetAnthropic | 67.4% | 20250219 | 50.2 | – |
| 29 | gpt-oss-20bOpenAIopen | 67.4% | – | 46.0 | – |
| 30 |  GLM-4.5-AirZhipu AIopen | 67.0% | – | 47.0 | – |
| 31 | DeepSeek V3DeepSeekopen | 66.5% | DeepSeek v3 | 45.2 | – |
| 32 | Gemini 1.5 Pro 002Google | 65.7% | Gemini 1.5 Pro (002) | 44.8 | – |
| 33 | GPT-5.1OpenAI | 65.6% | – | 59.4 | – |
| 34 | Claude 3.5 SonnetAnthropic | 65.3% | 20241022 | 46.3 | – |
| 35 | Llama 4 ScoutMetaopen | 64.4% | Llama 4 Scout (17Bx16E) Instruct | 40.4 | – |
| 36 | Gemini 2.0 Flash-LiteGoogle | 64.2% | – | 47.7 | – |
| 37 |  Nova PremierAmazon | 63.7% | – | 45.8 | – |
| 38 | GPT-4oOpenAI | 63.4% | – | 43.8 | – |
| 39 | Gemini 2.5 FlashGoogle | 62.6% | – | 50.0 | – |
| 40 | Llama 3.1 405BMetaopen | 61.8% | Llama 3.1 Instruct Turbo (405B) | 43.1 | – |
| 41 | GPT-4.1 nanoOpenAI | 61.6% | – | 39.1 | – |
| 42 | Gemini 1.5 Flash 002Google | 60.9% | Gemini 1.5 Flash (002) | 40.8 | – |
| 43 | Palmyra X 004Writer | 60.9% | – | – | – |
| 44 | Qwen2.5 72BAlibabaopen | 59.9% | Qwen2.5 Instruct Turbo (72B) | 42.7 | – |
| 45 |  Mistral Large 2Mistral AIopen | 59.8% | 2411 | 40.8 | – |
| 46 | Gemini 2.5 Flash-LiteGoogle | 59.1% | – | 43.8 | – |
| 47 | Nova ProAmazon | 59.1% | – | 41.1 | – |
| 48 | Palmyra FinWriter | 57.7% | – | – | – |
| 49 |  Granite 4.0 SmallIBM | 57.5% | IBM Granite 4.0 Small | – | – |
| 50 | Llama 3.1 70BMetaopen | 57.4% | Llama 3.1 Instruct Turbo (70B) | 41.1 | – |
| 51 | GPT-4o miniOpenAI | 56.5% | – | 36.6 | – |
| 52 | Mistral Small 3.1Mistral AIopen | 55.8% | 2503 | 40.3 | – |
| 53 | Amazon Nova LiteAmazon | 55.1% | – | 39.4 | – |
| 54 | Claude 3.5 HaikuAnthropic | 54.9% | 20241022 | 40.0 | – |
| 55 | Qwen2.5 7B InstructAlibabaopen | 52.9% | Qwen2.5 Instruct Turbo (7B) | 40.1 | – |
| 56 | Nova MicroAmazon | 52.2% | – | 39.8 | – |
| 57 | granite-4.0-microIBMopen | 48.6% | IBM Granite 4.0 Micro | 35.9 | – |
| 58 | Mixtral 8x22B InstructMistral AIopen | 47.8% | Mixtral Instruct (8x22B) | 37.6 | – |
| 59 | Palmyra MedWriter | 47.6% | – | – | – |
| 60 |  Olmo 2 32B March 2025Ai2 | 47.5% | OLMo 2 32B Instruct March 2025 | – | – |
Cite as: BenchLeader, “HELM Capabilities mean leaderboard”, https://www.benchleader.com/benchmarks/helm_mean, data as of 12 Sept 2026.
HELM Capabilities mean: questions
- What does HELM Capabilities mean measure?
- The mean of HELM's five capability scenarios: MMLU-Pro, GPQA, IFEval, WildBench and Omni-MATH. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads HELM Capabilities mean?
- GPT-5 mini leads HELM Capabilities mean with 81.9% as of 12 Sept 2026, ahead of o4-mini at 81.2%.
- How many models have HELM Capabilities mean results?
- 68 model configurations have a HELM Capabilities mean result on BenchLeader, all taken from HELM Capabilities.
- Who runs HELM Capabilities mean and how often is it updated?
- HELM Capabilities mean is published by HELM Capabilities. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does HELM Capabilities mean count toward the BenchLeader Index?
- No. HELM Capabilities mean is shown for reference but left out of the composite index.