BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

Arena Hard Prompts

Text arena rating on prompts judged hard.

As of 24 Oct 2026, Gemini 4 Argon leads Arena Hard Prompts on BenchLeader with 1540, ahead of Claude Opus 5.5 at 1535, across 373 model configurations with a published result.

Published by
Arena (LMArena)
Category
Reasoning
Index weight
0.5
Models
373
Data as of
24 Oct 2026
Newest result
22 Oct 2026

CC BY 4.0 — lmarena-ai/leaderboard-dataset on Hugging Face.

What the test looks like

The same pairwise votes as the text arena, restricted to prompts that LMArena's classifier judges hard: specific, complex, multi-step requests rather than casual chat.

How it is scored

Bradley-Terry rating on the hard-prompt subset, with style control.

What to keep in mind

Fewer votes per model than the overall board, so confidence intervals are wider. Hardness is judged by a classifier, not by outcome.

Top 15
  1. 1Gemini 4 Argon (high)1540
  2. 2Claude Opus 5.5 (high)1535
  3. 3Claude Opus 4.6 (high)1530
  4. 4Claude Fable 5 (high)1529
  5. 5Claude Opus 4.61527
  6. 6Claude Opus 4.7 (high)1525
  7. 7Claude Fable 5.1 (max)1523
  8. 8Muse Wonder (max)1522
  9. 9Claude Opus 4.71519
  10. 10Gemini 3.8 Flash (high)1517
  11. 11Muse Spark 1.3 (max)1517
  12. 12Kimi K3 (max)1516
  13. 13Claude Opus 5 (high)1516
  14. 14Claude Opus 5 (max)1515
  15. 15Claude Opus 4.8 (high)1513
373 of 373
#
1Gemini 4 Argonhigh154070.82026-10-22
2Claude Opus 5.5high153569.42026-10-22
3Claude Opus 4.6high153062.72026-10-22
4Claude Fable 5high152964.62026-10-22
5Claude Opus 4.6152762.62026-10-22
6Claude Opus 4.7high152561.02026-10-22
7Claude Fable 5.1max152370.52026-10-22
8Muse Wondermax152267.52026-10-22
9Claude Opus 4.7151963.32026-10-22
10Gemini 3.8 Flashhigh151764.42026-10-22
11Muse Spark 1.3max151767.12026-10-22
12Kimi K3max151664.72026-10-22
13Claude Opus 5high151667.92026-10-22
14Claude Opus 5max151568.82026-10-22
15Claude Opus 4.8high151360.72026-10-22
16GPT-5.6 Solxhigh151266.62026-10-22
17Muse Spark 1.2xhigh151263.22026-10-22
18Muse Spark 1.1151062.52026-10-22
19Gemini 3.7 Flashhigh150963.52026-10-22
20Gemini 3.1 Pro150860.62026-10-22
21Muse Spark150664.32026-10-22
22Claude Opus 4.8150560.12026-10-22
23Claude Sonnet 5.5xhigh150567.92026-10-22
24GLM 5.3max150563.52026-10-22
25GPT-6.1 Solmax150469.92026-10-22
26Qwen3.8 Maxmax150463.32026-10-22
27MiMo-V2.6-Pro150464.02026-10-22
28Claude Sonnet 4.6150357.12026-10-22
29Gemini 3 Pro150260.42026-10-22
30GPT-5.5high150165.32026-10-22
31Gemini 3.6 Flashhigh150162.02026-10-22
32GPT-6 Astramax150170.42026-10-22
33GLM 5.3 Flash150162.62026-10-22
34Claude Opus 4.5high150159.22026-10-22
35GPT-5.5149960.32026-10-22
36DeepSeek V4.1 Flashmax149961.22026-10-22
37Claude Opus 4.5149857.92026-10-22
38GPT-5.2149758.02026-10-22
39GPT-5.4high149757.92026-10-22
40GLM 5.2max149662.32026-10-22
41Gemini 3.5 Flashmedium149562.82026-10-22
42Qwen3.7max149560.62026-10-22
43Gemini 3.5 Flashhigh149563.12026-10-22
44MiMo-V2.5-Pro149454.92026-10-22
45Gemini 3 Flash149357.02026-10-22
46GPT-5.5 Instant149158.02026-10-22
47Qwen3.5 Maxmax1491–2026-10-22
48GPT-5.6 Terraxhigh149062.82026-10-22
49Claude Sonnet 5high149060.22026-10-22
50Grok 4.20thinking148958.62026-10-22
51GLM 5.1148955.92026-10-22
52Grok 4.5148859.02026-10-22
53Grok 4.20 11487–2026-10-22
54ERNIE 5.11487–2026-10-22
55GPT-5.4148758.22026-10-22
56Claude Sonnet 4.5high148758.12026-10-22
57Kimi K2.6148555.42026-10-22
58GPT-6 Solmax148566.42026-10-22
59Grok 4.20 Multi-Agent148458.52026-10-22
60Claude Sonnet 4.5148353.72026-10-22

More reasoning benchmarks · All LLM benchmarks

Cite as: BenchLeader, “Arena Hard Prompts leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/lmarena_text_hard.

Arena Hard Prompts: questions

What does Arena Hard Prompts measure?
The same pairwise votes as the text arena, restricted to prompts that LMArena's classifier judges hard: specific, complex, multi-step requests rather than casual chat. Scores are reported in Elo-style ratings from pairwise comparisons; higher is better.
Which AI model leads Arena Hard Prompts?
Gemini 4 Argon leads Arena Hard Prompts with 1540 as of 24 Oct 2026, ahead of Claude Opus 5.5 at 1535.
How many models have Arena Hard Prompts results?
373 model configurations have a Arena Hard Prompts result on BenchLeader, all taken from Arena (LMArena).
Who runs Arena Hard Prompts and how often is it updated?
Arena Hard Prompts is published by Arena (LMArena). BenchLeader re-reads the published results once a day and records the date each result was published.
Does Arena Hard Prompts count toward the BenchLeader Index?
Yes. Arena Hard Prompts contributes to the reasoning category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.