HLE w/o tools leaderboard

Tool-free variant of Humanity's Last Exam that isolates a model's raw frontier reasoning.

Models ranked by their published Humanity's Last Exam without tools value

Rank	Model	Provider	Tool-free expert QA
1	Claude Fable 5.1	Anthropic	60.9
2	Claude Mythos 5	Anthropic	59
3	Claude Opus 5	Anthropic	56.3
4	Muse Spark 1.1	Meta	52.2
5	SASakana Fugu-Ultra	Sakana AI	50
6	Claude Opus 4.8	Anthropic	49.8
7	UNPareto 26.9	Unbiased	49
8	SASakana Fugu	Sakana AI	47.2
9	Claude Opus 4.7 (Adaptive)	Anthropic	46.9
10	Gemini 3.1 Pro	Google	45.4
11	OAOrnith-1.5-397B	Ornith AI	44.6
12	Qwen3.8 Max	Alibaba	43.6
13	Kimi K3	Moonshot AI	43.5
14	Hy4 preview	Tencent	43.4
15	Claude Sonnet 5	Anthropic	43.2
16	GPT-5.5 Pro	OpenAI	43.1
17	Muse Spark	Meta	42.8
18	GPT-5.4 Pro	OpenAI	42.7
19	GPT-5.5	OpenAI	41.4
20	GLM-5.2	Z.AI	40.5
21	Claude Opus 4.6	Anthropic	40
22	GPT-5.4	OpenAI	39.8
23	Qwen3.8-Flash-Next	Alibaba	35.9
24	MiMo-V2.5-Pro	Xiaomi	34
25	Grok 4.20	xAI	31.6
25	TMInkling-Small	Thinking Machines Lab	31.6
27	Qwen3.8-27B	Alibaba	30.8
28	TMInkling	Thinking Machines Lab	30
29	Solar Open 2	Upstage	28.8
30	GPT-5.4 mini	OpenAI	28.2
31	Nemotron 3 Ultra	NVIDIA	26.7
32	OAOrnith-1.5-35B-A3B	Ornith AI	25.6
33	GPT-5.4 nano	OpenAI	24.3
34	OAOrnith-1.5-9B	Ornith AI	20.2
35	Gemma 4 31B	Google	19.5
36	Nemotron 3.5 Lightning 30B A3B NVFP4	NVIDIA	10.47
37	Gemma 4 26B A4B	Google	8.7
38	Gemma 4 12B	Google	5.2

Evidence key: Observed

Rows are ordered by the value Introducing GPT-5.4 mini and nano published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.