Live ranking, refreshed every 2 days
Best AI Models for Hard Prompts
As of October 8, 2026, the top-ranked model on LMArena's text leaderboard, hard prompts category is gemini-4-argon-high by Google with a score of 1,551, 18 points ahead of claude-opus-5.5-high by Anthropic.
Updated October 13, 2026. Showing the last saved copy of the leaderboard, dated October 8, 2026, because the live source could not be reached.
Ranking: top 10 models
Scores come from blind head-to-head votes on LMArena's text leaderboard, hard prompts category. Rank is LMArena's published rank, which accounts for confidence intervals, so several models can share a rank. The top 10 are shown below, and ranks 11 to 40 follow in the expandable list.
| Rank | Model | Lab | Score | Votes | License |
|---|---|---|---|---|---|
| 1 | gemini-4-argon-high | 1,551 ±11 | 3,118 | Proprietary | |
| 2 | claude-opus-5.5-high | Anthropic | 1,533 ±10 | 3,901 | Proprietary |
| 3 | claude-opus-4-6-high | Anthropic | 1,533 ±4 | 50,602 | Proprietary |
| 4 | claude-fable-5-high | Anthropic | 1,530 ±5 | 26,112 | Proprietary |
| 5 | claude-opus-4-6 | Anthropic | 1,527 ±4 | 54,602 | Proprietary |
| 6 | claude-opus-4-7-high | Anthropic | 1,525 ±5 | 43,711 | Proprietary |
| 7 | claude-fable-5.1-max | Anthropic | 1,520 ±8 | 7,642 | Proprietary |
| 8 | claude-opus-4-7 | Anthropic | 1,519 ±5 | 44,480 | Proprietary |
| 9 | gemini-3.8-flash-high | 1,517 ±6 | 20,342 | Proprietary | |
| 10 | kimi-k3-max | Moonshot AI | 1,517 ±5 | 19,507 | Open weights |
Show ranks 11 to 40
Rank Model Lab Score License 11 muse-spark-1.3-max Meta 1,517 ±7 Proprietary 12 claude-opus-5-high Anthropic 1,515 ±5 Proprietary 13 claude-opus-5-max Anthropic 1,514 ±6 Proprietary 14 claude-opus-4-8-high Anthropic 1,513 ±5 Proprietary 15 gpt-5.6-sol-xhigh OpenAI 1,512 ±5 Proprietary 16 muse-spark-1.1 Meta 1,511 ±5 Proprietary 17 gemini-3.1-pro-preview Google 1,508 ±4 Proprietary 18 muse-spark-1.2 (xHigh) Meta 1,508 ±10 Proprietary 19 claude-sonnet-5.5-xhigh Anthropic 1,507 ±11 Proprietary 20 gpt-6.1-sol-max OpenAI 1,507 ±12 Proprietary 21 gemini-3.7-flash-high Google 1,506 ±6 Proprietary 22 muse-spark Meta 1,506 ±7 Proprietary 23 mimo-v2.6-pro Xiaomi 1,506 ±12 Open weights 24 claude-opus-4-8 Anthropic 1,505 ±5 Proprietary 25 qwen3.8-max Alibaba (Qwen) 1,504 ±6 Proprietary 26 glm-5.3-max Z.ai (Zhipu) 1,504 ±6 Open weights 27 claude-sonnet-4-6 Anthropic 1,503 ±4 Proprietary 28 gemini-3-pro Google 1,502 ±5 Proprietary 29 gpt-5.5-high OpenAI 1,501 ±4 Proprietary 30 glm-5.3-flash Z.ai (Zhipu) 1,501 ±6 Open weights 31 claude-opus-4-5-20251101-high-32k Anthropic 1,501 ±5 Proprietary 32 gemini-3.6-flash-high Google 1,501 ±5 Proprietary 33 gpt-6-astra-max OpenAI 1,500 ±8 Proprietary 34 deepseek-v4.1-flash-max DeepSeek 1,499 ±8 Open weights 35 gpt-5.5 OpenAI 1,499 ±4 Proprietary 36 claude-opus-4-5-20251101 Anthropic 1,498 ±4 Proprietary 37 gpt-5.2-chat-latest-20260210 OpenAI 1,497 ±5 Proprietary 38 gpt-5.4-high OpenAI 1,496 ±5 Proprietary 39 mimo-v2.5-pro Xiaomi 1,496 ±4 Open weights 40 qwen3.7-max-preview Alibaba (Qwen) 1,496 ±12 Proprietary
Showing the top 40 of 414 models on this leaderboard. Data from October 8, 2026.
Best model by lab
Each lab's highest-ranked model on this leaderboard, and how far it sits from the top.
| Lab | Best model | Rank | Score | Behind #1 |
|---|---|---|---|---|
| gemini-4-argon-high | 1 | 1,551 | Leader | |
| Anthropic | claude-opus-5.5-high | 2 | 1,533 | 18 pts |
| Moonshot AI | kimi-k3-max | 10 | 1,517 | 34 pts |
| Meta | muse-spark-1.3-max | 11 | 1,517 | 34 pts |
| OpenAI | gpt-5.6-sol-xhigh | 15 | 1,512 | 39 pts |
| Xiaomi | mimo-v2.6-pro | 23 | 1,506 | 45 pts |
| Alibaba (Qwen) | qwen3.8-max | 25 | 1,504 | 47 pts |
| Z.ai (Zhipu) | glm-5.3-max | 26 | 1,504 | 47 pts |
Which AI lab leads?
The lab with the most models at the top of the board is the one with the deepest bench, not only the single best model.
Anthropic holds 7 of the top 10 models (70%).
Models in the top 10 by lab
Show data table
| Lab | Top 10 |
|---|---|
| Anthropic | 7 |
| 2 | |
| Moonshot AI | 1 |
Open vs proprietary: how big is the gap?
The best open-weight model is kimi-k3-max by Moonshot AI at 1,517, 34 points behind the best proprietary model, gemini-4-argon-high by Google at 1,551. Their 95% confidence intervals do not overlap, so the gap is statistically meaningful.
9 proprietary models rank above the best open-weight model, and 1 of the top 20 is open-weight.
Open-weight here means any license other than Proprietary, including licenses with usage limits. It is a broader definition than the OSI one.
Other leaderboards
LMArena runs separate arenas for different jobs. The best model for one is often not the best for another.
Other arenas
- Text (overall)
- WebDev: claude-opus-5.5-max by Anthropic, 1,813
- Vision: claude-fable-5-high by Anthropic, 1,308
- Search: gpt-5.6-sol-xhigh by OpenAI, 1,256
- Document: claude-fable-5 by Anthropic, 1,499
- Text to image: gpt-image-2.5-sunburst by OpenAI, 1,425
- Image editing: gpt-image-2.5-sunburst by OpenAI, 1,524
- Text to video: gemini-omni-1.1-flash by Google, 1,516
Text arena by category
How to read this leaderboard
This category keeps the prompts LMArena classifies as hard: complex, specific and demanding. It separates models that handle everyday questions from those that handle difficult ones.
- Scores come from blind votes
- LMArena shows a person two anonymous answers to the same prompt and asks which is better. Those head-to-head votes are fitted with a Bradley-Terry model, the statistical successor to Elo, to give each model one arena score. On the Elo scale LMArena uses, a 100-point lead means the higher model is expected to win about 64% of matchups.
- Confidence intervals decide ties
- Every score carries a 95% confidence interval. When two intervals overlap, the order of those two models is not settled. LMArena's rank column already accounts for this, which is why more than one model can hold the same rank.
- Why it differs from usage share
- A high rank does not mean many people use a model. Usage depends on price, distribution and habit. For adoption rather than quality, see AI market share.
Methodology and source
All rankings are read from the LMArena leaderboard dataset on Hugging Face, published by LMArena (formerly Chatbot Arena) under the Creative Commons Attribution 4.0 license. Credit for the underlying votes, scores and confidence intervals belongs to LMArena. Essentras is not affiliated with LMArena.
This page reads the text_style_control subset, category hard_prompts, from the dataset's latest split. Scores, intervals, ranks, vote counts and license labels are shown as published. Essentras adds the grouping by lab and the derived figures on this page: gaps, lab shares, open vs proprietary and, on the main page, the history of who held #1.
Data refreshes automatically every 2 days. If the source cannot be reached, the page falls back to the last saved copy and says so at the top. Models are never merged or renamed; model and lab names appear exactly as LMArena publishes them.
Open-weight means any license other than Proprietary. History is limited to what the dataset records: the style-controlled text leaderboard from August 28, 2024. We do not scrape any other ranking site.
Track how AI models rank and describe your brand with our AI SEO service. Want these models to name your brand? Read our guide to LLM SEO.
Frequently asked questions
What is the best AI model for hard prompts?
As of October 8, 2026, gemini-4-argon-high by Google ranks first on LMArena's text leaderboard, hard prompts category with a score of 1,551. Rankings move as models and votes are added, so check the table above for the current order.
Which AI lab leads for hard prompts?
Anthropic holds 7 of the top 10 models (70%), ahead of Google (2) and Moonshot AI (1).
Is the best open-source model close to the best proprietary model?
The best open-weight model is kimi-k3-max by Moonshot AI at 1,517, 34 points behind the best proprietary model, gemini-4-argon-high by Google at 1,551. 9 proprietary models rank above the best open-weight model, and 1 of the top 20 is open-weight. Their 95% confidence intervals do not overlap, so the gap is statistically meaningful. "Open" here means any license other than Proprietary, which is broader than the OSI definition.
What does this leaderboard measure?
This category keeps the prompts LMArena classifies as hard: complex, specific and demanding. It separates models that handle everyday questions from those that handle difficult ones.
How often does this leaderboard update?
This page re-reads LMArena's dataset every 2 days. LMArena publishes a new snapshot when it has gathered enough votes or added models, so the date at the top shows the latest published data.
Models rank themselves. What do they say about you?
A leaderboard shows which AI models are strongest. It does not show what those models tell your customers about your company. Run the free check and see the answers.