LLM Rankings by Category: The Hidden Specialists
22 of the Arena's top 60 models rank at least 10 places higher in one category than overall. The overall list hides specialists.
The signal
- LaunchesGPT-5.6 Sol, Luna and Terra (9 Jul 2026). On 13 Sept 2026, 4 OpenAI models in the overall top 60 rank 10+ places higher on expert prompts.
- ArenaBest LLM for hard prompts on 13 Sept 2026: claude-opus-4-6-high (Anthropic), #4 overall.
- ArenaModels that gain 10+ places, by their strongest category: expert 6, coding 5, creative writing 5, longer queries 3, math 1, hard prompts 1, instruction following 1.
claude-opus-4-5-20251101-high-32k is #49 overall but #14 in coding. gpt-5.6-terra-xhigh is #52 overall but #24 in expert prompts.
Places gained in the model's strongest Arena category vs its overall rank, top 60 overall models, 13 Sept 2026. One bar per model; top 14 shown.
Categories: coding, math, creative writing, hard prompts, instruction following, longer queries, expert. Arena category boards re-score the same blind votes, filtered by prompt type. For each category we rank the overall and category boards over the models present on both, so a smaller board does not move ranks by itself. Only category ratings with at least 1,000 votes count. Neighbouring ranks often overlap within confidence intervals (±5–14 points here), so read gains under about 10 places as noise.
Show the data (14 rows)
| Model | Lab | Category | Overall rank | Category rank | Places gained | Votes |
|---|---|---|---|---|---|---|
| claude-opus-4-5-20251101-high-32k | Anthropic | coding | 49 | 14 | 35 | 7,515 |
| gpt-5.6-terra-xhigh | OpenAI | expert | 52 | 24 | 28 | 3,169 |
| claude-opus-4-5-20251101 | Anthropic | coding | 44 | 19 | 25 | 17,174 |
| gpt-5.6-sol-xhigh | OpenAI | expert | 37 | 14 | 23 | 2,954 |
| claude-sonnet-4-6 | Anthropic | coding | 35 | 13 | 22 | 18,320 |
| claude-opus-4-8 | Anthropic | longer queries | 40 | 22 | 18 | 24,899 |
| mimo-v2.5-pro | Xiaomi | coding | 32 | 16 | 16 | 16,683 |
| kimi-k2.5-thinking | Moonshot | math | 51 | 36 | 15 | 3,916 |
| claude-opus-4-8-high | Anthropic | longer queries | 34 | 20 | 14 | 24,417 |
| deepseek-v4-pro-high-preview | DeepSeek | creative writing | 54 | 40 | 14 | 8,621 |
| glm-5.3-flash | Z.ai | coding | 24 | 11 | 13 | 2,755 |
| gpt-5.5 | OpenAI | expert | 31 | 18 | 13 | 6,976 |
| gemini-2.5-pro | creative writing | 36 | 23 | 13 | 17,131 | |
| kimi-k3-max | Moonshot | expert | 23 | 11 | 12 | 1,863 |
The Mode read
Most leaderboard readers stop at the overall list. But the Arena also scores each model on categories of prompts: coding, math, expert questions, long queries and more. Is the overall rank a good guide to which model fits a given job?
Often it is not. 22 of the top 60 models rank at least 10 places higher in one category. Anthropic's claude-opus-4-5-20251101-high-32k, a release dated Nov 2025 in its name, is #49 overall but #14 in coding. OpenAI's gpt-5.6-terra-xhigh is #52 overall and #24 on expert prompts. Each of these gains rests on at least 1,000 category votes.
The overall rank is an average over everyone's prompts. A vertical product serves one kind of prompt, so it should test on its own work, not on the average. Teams that test this way can pick a model that most buyers overlook. The work of testing also builds a set of real customer tasks that no competitor has.
Overall rank picks a model for everyone. Your customers are not everyone.
What we would back: Vertical products that score models on their own customers' tasks and keep that test set as an asset.