Which AI model leads WildBench (HELM)?

Qwen3 235B A22B 2507 leads WildBench (HELM) with 86.6% as of 12 Sept 2026, ahead of GPT-5.1 at 86.3%.

How many models have WildBench (HELM) results?

68 model configurations have a WildBench (HELM) result on BenchLeader, all taken from HELM Capabilities.

Who runs WildBench (HELM) and how often is it updated?

WildBench (HELM) is published by HELM Capabilities. BenchLeader re-reads the published results every morning and records the date each result was published.

Does WildBench (HELM) count toward the BenchLeader Index?

Yes. WildBench (HELM) contributes to the instruction following category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.

WildBench (HELM)

Real user tasks graded by a rubric-guided judge. Run by Stanford CRFM's HELM.

As of 12 Sept 2026, Qwen3 235B A22B 2507 leads WildBench (HELM) on BenchLeader with 86.6%, ahead of GPT-5.1 at 86.3%, across 68 model configurations with a published result.

Published by: HELM Capabilities
Category: Instruction following
Index weight: 0.5
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

WildBench draws real tasks from users' conversations with chatbots, from coding help to creative writing, and grades responses with a judge model following task-specific checklists.

How it is scored

WB Score, a 0 to 100 rating from the checklist-guided judge, run by Stanford CRFM's HELM.

What to keep in mind

Judged by another model, so it inherits that judge's preferences; BenchLeader gives it half weight for that reason.

Top 15

1Qwen3 235B A22B 250786.6%
2GPT-5.186.3%
3Kimi K286.2%
4o386.1%
5Gemini 3 Pro85.9%
6GPT-585.7%
7Gemini 2.5 Pro85.7%
8GPT-5 mini85.5%
9o4-mini85.4%
10Claude Sonnet 4.585.4%
11GPT-4.185.4%
12Claude Opus 485.2%
13Grok 384.9%
14gpt-oss-120b84.5%
15Claude Haiku 4.583.9%

68 of 68

#
1	Qwen3 235B A22B 2507Alibabaopen	86.6%	Qwen3 235B A22B Instruct 2507 FP8	54.1	–
2	GPT-5.1OpenAI	86.3%	–	59.4	–
3	Kimi K2Moonshot AIopen	86.2%	Kimi K2 Instruct	51.2	–
4	o3OpenAI	86.1%	–	61.6	–
5	Gemini 3 ProGoogle	85.9%	–	62.0	–
6	GPT-5OpenAI	85.7%	–	60.8	–
7	Gemini 2.5 ProGoogle	85.7%	–	54.5	–
8	GPT-5 miniOpenAI	85.5%	–	57.9	–
9	o4-miniOpenAI	85.4%	–	57.2	–
10	Claude Sonnet 4.5Anthropic	85.4%	20250929	53.7	–
11	GPT-4.1OpenAI	85.4%	–	48.8	–
12	Claude Opus 4Anthropic	85.2%	–	56.1	–
13	Grok 3xAI	84.9%	–	51.0	–
14	gpt-oss-120bOpenAIopen	84.5%	–	48.8	–
15	Claude Haiku 4.5Anthropic	83.9%	20251001	48.6	–
16	Claude Sonnet 4Anthropic	83.8%	–	55.1	–
17	GPT-4.1 miniOpenAI	83.8%	–	46.3	–
18	Claude Opus 4Anthropic	83.3%	20250514	52.2	–
19	DeepSeek V3DeepSeekopen	83.1%	DeepSeek v3	45.2	–
20	DeepSeek R1 0528DeepSeekopen	82.8%	–	49.9	–
21	Qwen3 235B A22BAlibabaopen	82.8%	Qwen3 235B A22B FP8 Throughput	44.8	–
22	GPT-4oOpenAI	82.8%	–	43.8	–
23	Claude Sonnet 4Anthropic	82.5%	20250514	50.5	–
24	Gemini 2.5 Flash-LiteGoogle	81.8%	–	43.8	–
25	Gemini 2.5 FlashGoogle	81.7%	–	50.0	–
26	Claude 3.7 SonnetAnthropic	81.4%	20250219	50.2	–
27	Gemini 1.5 Pro 002Google	81.3%	Gemini 1.5 Pro (002)	44.8	–
28	GPT-4.1 nanoOpenAI	81.1%	–	39.1	–
29	Qwen3-Next 80B A3BAlibabaopen	80.7%	–	51.2	–
30	GPT-5 nanoOpenAI	80.6%	–	52.5	–
31	Qwen2.5 72BAlibabaopen	80.2%	Qwen2.5 Instruct Turbo (72B)	42.7	–
32	Palmyra X 004Writer	80.2%	–	–	–
33	Mistral Large 2Mistral AIopen	80.1%	2411	40.8	–
34	Llama 4 MaverickMetaopen	80.0%	Llama 4 Maverick (17Bx128E) Instruct FP8	44.2	–
35	Gemini 2.0 FlashGoogle	80.0%	Gemini 2.0 Flash	43.5	–
36	Grok 4xAI	79.7%	0709	58.6	–
37	Claude 3.5 SonnetAnthropic	79.2%	20241022	46.3	–
38	Gemini 1.5 Flash 002Google	79.2%	Gemini 1.5 Flash (002)	40.8	–
39	GPT-4o miniOpenAI	79.1%	–	36.6	–
40	Gemini 2.0 Flash-LiteGoogle	79.0%	–	47.7	–
41	GLM-4.5-AirZhipu AIopen	78.9%	–	47.0	–
42	Nova PremierAmazon	78.8%	–	45.8	–
43	Mistral Small 3.1Mistral AIopen	78.8%	2503	40.3	–
44	Llama 3.1 405BMetaopen	78.3%	Llama 3.1 Instruct Turbo (405B)	43.1	–
45	Palmyra FinWriter	78.3%	–	–	–
46	Palmyra X5Writer	78.0%	Palmyra X5	–	–
47	Llama 4 ScoutMetaopen	77.9%	Llama 4 Scout (17Bx16E) Instruct	40.4	–
48	Nova ProAmazon	77.7%	–	41.1	–
49	Claude 3.5 HaikuAnthropic	76.0%	20241022	40.0	–
50	Llama 3.1 70BMetaopen	75.8%	Llama 3.1 Instruct Turbo (70B)	41.1	–
51	Amazon Nova LiteAmazon	75.0%	–	39.4	–
52	Nova MicroAmazon	74.3%	–	39.8	–
53	Granite 3.3 8BIBM	74.1%	IBM Granite 3.3 8B Instruct	–	–
54	Granite 4.0 SmallIBM	73.9%	IBM Granite 4.0 Small	–	–
55	gpt-oss-20bOpenAIopen	73.7%	–	46.0	–
56	Olmo 2 32B March 2025Ai2	73.4%	OLMo 2 32B Instruct March 2025	–	–
57	Qwen2.5 7B InstructAlibabaopen	73.1%	Qwen2.5 Instruct Turbo (7B)	40.1	–
58	Mixtral 8x22B InstructMistral AIopen	71.1%	Mixtral Instruct (8x22B)	37.6	–
59	Olmo 2 13B November 2024Ai2	68.9%	OLMo 2 13B Instruct November 2024	–	–
60	Llama 3.1 8BMetaopen	68.6%	Llama 3.1 Instruct Turbo (8B)	36.3	–

Cite as: BenchLeader, “WildBench (HELM) leaderboard”, https://www.benchleader.com/benchmarks/helm_wildbench, data as of 12 Sept 2026.

WildBench (HELM): questions

What does WildBench (HELM) measure?: WildBench draws real tasks from users' conversations with chatbots, from coding help to creative writing, and grades responses with a judge model following task-specific checklists. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads WildBench (HELM)?: Qwen3 235B A22B 2507 leads WildBench (HELM) with 86.6% as of 12 Sept 2026, ahead of GPT-5.1 at 86.3%.
How many models have WildBench (HELM) results?: 68 model configurations have a WildBench (HELM) result on BenchLeader, all taken from HELM Capabilities.
Who runs WildBench (HELM) and how often is it updated?: WildBench (HELM) is published by HELM Capabilities. BenchLeader re-reads the published results every morning and records the date each result was published.
Does WildBench (HELM) count toward the BenchLeader Index?: Yes. WildBench (HELM) contributes to the instruction following category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.