IFEval (HELM)
Strict instruction following on verifiable constraints. Run by Stanford CRFM's HELM.
As of 12 Sept 2026, Grok 3 mini leads IFEval (HELM) on BenchLeader with 95.1%, ahead of Grok 4 at 94.9%, across 68 model configurations with a published result.
- Published by
- HELM Capabilities
- Category
- Instruction following
- Index weight
- 1.0
- 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
IFEval gives the model writing tasks with verifiable instructions attached: respond in exactly three paragraphs, avoid the letter e, end with a given phrase, and so on. Each instruction can be checked by a program, so no judge model is needed.
How it is scored
Strict accuracy: the share of instructions satisfied exactly, as run by Stanford CRFM's HELM in its fixed harness.
What to keep in mind
It checks whether rules were obeyed, not whether the writing was any good, and the constraints are artificial. Near saturation for frontier models, so it mainly separates the mid-field.
- 1Grok 3 mini95.1%
- 2Grok 494.9%
- 3GPT-5.193.5%
- 4GPT-5 nano93.2%
- 5o4-mini92.9%
- 6GPT-5 mini92.7%
- 7Claude Opus 491.8%
- 8Llama 4 Maverick90.8%
- 9GPT-4.1 mini90.4%
- 10Gemini 2.5 Flash89.8%
- 11Granite 4.0 Small89.0%
- 12Grok 388.4%
- 13Gemini 3 Pro87.6%
- 14Mistral Large 287.6%
- 15GPT-587.5%
| # | |||||
|---|---|---|---|---|---|
| 1 |  Grok 3 minixAI | 95.1% | – | 51.8 | – |
| 2 | Grok 4xAI | 94.9% | 0709 | 58.6 | – |
| 3 |  GPT-5.1OpenAI | 93.5% | – | 59.4 | – |
| 4 | GPT-5 nanoOpenAI | 93.2% | – | 52.5 | – |
| 5 | o4-miniOpenAI | 92.9% | – | 57.2 | – |
| 6 | GPT-5 miniOpenAI | 92.7% | – | 57.9 | – |
| 7 |  Claude Opus 4Anthropic | 91.8% | 20250514 | 52.2 | – |
| 8 |  Llama 4 MaverickMetaopen | 90.8% | Llama 4 Maverick (17Bx128E) Instruct FP8 | 44.2 | – |
| 9 | GPT-4.1 miniOpenAI | 90.4% | – | 46.3 | – |
| 10 |  Gemini 2.5 FlashGoogle | 89.8% | – | 50.0 | – |
| 11 |  Granite 4.0 SmallIBM | 89.0% | IBM Granite 4.0 Small | – | – |
| 12 | Grok 3xAI | 88.4% | – | 51.0 | – |
| 13 | Gemini 3 ProGoogle | 87.6% | – | 62.0 | – |
| 14 |  Mistral Large 2Mistral AIopen | 87.6% | 2411 | 40.8 | – |
| 15 | GPT-5OpenAI | 87.5% | – | 60.8 | – |
| 16 |  Palmyra X 004Writer | 87.2% | – | – | – |
| 17 | o3OpenAI | 86.9% | – | 61.6 | – |
| 18 | Claude 3.5 SonnetAnthropic | 85.6% | 20241022 | 46.3 | – |
| 19 | Claude Sonnet 4.5Anthropic | 85.0% | 20250929 | 53.7 | – |
| 20 |  Kimi K2Moonshot AIopen | 85.0% | Kimi K2 Instruct | 51.2 | – |
| 21 | Claude Opus 4Anthropic | 84.9% | – | 56.1 | – |
| 22 | granite-4.0-microIBMopen | 84.9% | IBM Granite 4.0 Micro | 35.9 | – |
| 23 | GPT-4.1 nanoOpenAI | 84.3% | – | 39.1 | – |
| 24 | Gemini 2.0 FlashGoogle | 84.1% | Gemini 2.0 Flash | 43.5 | – |
| 25 | Claude Sonnet 4Anthropic | 84.0% | – | 55.1 | – |
| 26 | Gemini 2.5 ProGoogle | 84.0% | – | 54.5 | – |
| 27 | Claude Sonnet 4Anthropic | 83.9% | 20250514 | 50.5 | – |
| 28 | GPT-4.1OpenAI | 83.8% | – | 48.8 | – |
| 29 | Gemini 1.5 Pro 002Google | 83.7% | Gemini 1.5 Pro (002) | 44.8 | – |
| 30 | gpt-oss-120bOpenAIopen | 83.6% | – | 48.8 | – |
| 31 |  Qwen3 235B A22B 2507Alibabaopen | 83.5% | Qwen3 235B A22B Instruct 2507 FP8 | 54.1 | – |
| 32 | Claude 3.7 SonnetAnthropic | 83.4% | 20250219 | 50.2 | – |
| 33 |  DeepSeek V3DeepSeekopen | 83.2% | DeepSeek v3 | 45.2 | – |
| 34 | Gemini 1.5 Flash 002Google | 83.1% | Gemini 1.5 Flash (002) | 40.8 | – |
| 35 | Gemini 2.0 Flash-LiteGoogle | 82.4% | – | 47.7 | – |
| 36 | Palmyra X5Writer | 82.3% | Palmyra X5 | – | – |
| 37 | Llama 3.1 70BMetaopen | 82.1% | Llama 3.1 Instruct Turbo (70B) | 41.1 | – |
| 38 | Llama 4 ScoutMetaopen | 81.8% | Llama 4 Scout (17Bx16E) Instruct | 40.4 | – |
| 39 | GPT-4oOpenAI | 81.7% | – | 43.8 | – |
| 40 | Qwen3 235B A22BAlibabaopen | 81.6% | Qwen3 235B A22B FP8 Throughput | 44.8 | – |
| 41 |  Nova ProAmazon | 81.5% | – | 41.1 | – |
| 42 |  GLM-4.5-AirZhipu AIopen | 81.2% | – | 47.0 | – |
| 43 | Llama 3.1 405BMetaopen | 81.1% | Llama 3.1 Instruct Turbo (405B) | 43.1 | – |
| 44 | Qwen3-Next 80B A3BAlibabaopen | 81.0% | – | 51.2 | – |
| 45 | Gemini 2.5 Flash-LiteGoogle | 81.0% | – | 43.8 | – |
| 46 | Qwen2.5 72BAlibabaopen | 80.6% | Qwen2.5 Instruct Turbo (72B) | 42.7 | – |
| 47 | Nova PremierAmazon | 80.3% | – | 45.8 | – |
| 48 | Claude Haiku 4.5Anthropic | 80.1% | 20251001 | 48.6 | – |
| 49 | Palmyra FinWriter | 79.3% | – | – | – |
| 50 | Claude 3.5 HaikuAnthropic | 79.2% | 20241022 | 40.0 | – |
| 51 | DeepSeek R1 0528DeepSeekopen | 78.4% | – | 49.9 | – |
| 52 | GPT-4o miniOpenAI | 78.2% | – | 36.6 | – |
| 53 |  Olmo 2 32B March 2025Ai2 | 78.0% | OLMo 2 32B Instruct March 2025 | – | – |
| 54 | Amazon Nova LiteAmazon | 77.6% | – | 39.4 | – |
| 55 | Palmyra MedWriter | 76.7% | – | – | – |
| 56 | Nova MicroAmazon | 76.0% | – | 39.8 | – |
| 57 | Mistral Small 3.1Mistral AIopen | 75.0% | 2503 | 40.3 | – |
| 58 | Llama 3.1 8BMetaopen | 74.3% | Llama 3.1 Instruct Turbo (8B) | 36.3 | – |
| 59 | Qwen2.5 7B InstructAlibabaopen | 74.1% | Qwen2.5 Instruct Turbo (7B) | 40.1 | – |
| 60 | gpt-oss-20bOpenAIopen | 73.2% | – | 46.0 | – |
Cite as: BenchLeader, “IFEval (HELM) leaderboard”, https://www.benchleader.com/benchmarks/helm_ifeval, data as of 12 Sept 2026.
IFEval (HELM): questions
- What does IFEval (HELM) measure?
- IFEval gives the model writing tasks with verifiable instructions attached: respond in exactly three paragraphs, avoid the letter e, end with a given phrase, and so on. Each instruction can be checked by a program, so no judge model is needed. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads IFEval (HELM)?
- Grok 3 mini leads IFEval (HELM) with 95.1% as of 12 Sept 2026, ahead of Grok 4 at 94.9%.
- How many models have IFEval (HELM) results?
- 68 model configurations have a IFEval (HELM) result on BenchLeader, all taken from HELM Capabilities.
- Who runs IFEval (HELM) and how often is it updated?
- IFEval (HELM) is published by HELM Capabilities. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does IFEval (HELM) count toward the BenchLeader Index?
- Yes. IFEval (HELM) contributes to the instruction following category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.