LLM Leaderboards Roundup (Oct 2026): Who Leads LMArena, LiveBench, HLE and More
A roundup of the current LLM leaderboards as of 19 Oct 2026: LMArena, Artificial Analysis, LiveBench, Scale HLE, OpenRouter and more, with ranking tables, source links and notes on stale boards.
Short answer: there is no single "best LLM" right now. On 19 October 2026 (KST) the leader depends on the board: Google's Gemini 4 Argon (preliminary) tops LMArena text, Anthropic's Claude Opus 5.5 and Fable 5.1 lead Artificial Analysis, LiveBench and LMArena WebDev, and OpenAI's GPT-6 Astra leads Scale's Humanity's Last Exam. This roundup collects the main boards in one place, with each table's original link directly underneath it.
How to read this: we re-fetched the live pages for LMArena (text, WebDev, Vision), Artificial Analysis, LiveBench, OpenRouter and Scale's HLE page on 19 Oct 2026 (KST) and copied the numbers below. Boards marked "not re-verified" come from our earlier read of the same day and could not be re-fetched in this pass. Model names are written as each site shows them. We did not run any benchmark ourselves. Each board below pairs a bar chart with its table. The charts use only the numbers in the tables; Elo-style charts start their axis above zero, which is noted in each chart footer.
At a glance: who leads which board
| Board | #1 | Data date | Type |
|---|---|---|---|
| LMArena Text | Gemini 4 Argon (High), preliminary | Oct 15, 2026 | Human preference votes |
| LMArena WebDev | Claude Opus 5.5 (Max) | Oct 18, 2026 | Human preference votes |
| LMArena Vision | Claude Fable 5 (High) | Oct 9, 2026 | Human preference votes |
| Artificial Analysis Intelligence Index | Claude Opus 5.5 (max) | live, no date shown | Independent benchmark suite |
| LiveBench 2.0 | Claude Fable 5.1 | updated Oct 18, 2026 | Long-horizon task suite |
| Scale HLE (detail page) | GPT 6 Astra | live, HLE-Rolling note Sep 17, 2026 | Academic questions |
| Epoch ECI | Claude Opus 5.5 (167) | announced Oct 1, 2026 | Composite index (table unreadable) |
| OpenRouter | DeepSeek V4.1 Flash | usage through Oct 17, 2026 | Usage, not quality |
LMArena Text (human preference)
Data as of Oct 15, 2026: 8,834,211 votes, 416 models. Scores show the ± interval the site prints.
| Rank | Model | Score |
|---|---|---|
| 1 | Gemini 4 Argon (High), Preliminary | 1524±9 |
| 2 | Muse Wonder (Max) (Meta) | 1509±11 |
| 3 | Claude Opus 5.5 (High) | 1508±7 |
| 4 | Claude Fable 5 (High) | 1504±4 |
| 5 | Claude Opus 4.6 (High) | 1503±3 |
| 6 | Claude Fable 5.1 (Max) | 1502±6 |
| 7 | Claude Opus 4.7 (High) | 1501±4 |
| 8 | Claude Opus 4.6 | 1497±3 |
Best open or Chinese-lab models on this board: Kimi K3 (Max) #18 at 1487, Qwen3.8-Max #24 at 1482 (proprietary), MiMo-V2.6-Pro #27 at 1479 (MIT), GLM-5.3 (Max) #28 at 1479 (MIT), DeepSeek-V4.1-Flash-20260910 (Max) #38 at 1475 (MIT).
Sources: LMArena Text leaderboard
LMArena WebDev
Data as of Oct 18, 2026: 146 models.
| Rank | Model | Score |
|---|---|---|
| 1 | Claude Opus 5.5 (Max) | 1806 |
| 2 | GPT-6 Astra (Max) | 1776 |
| 3 | Claude Sonnet 5.5 (xHigh) | 1768 |
| 4 | GPT-6.1 Sol (Max) | 1754 |
| 5 | Claude Fable 5.1 (Max) | 1742 |
| 6 | Claude Sonnet 5.5 (High) | 1709 |
| 7 | Claude Opus 5 (Max) | 1690 |
| 8 | GPT-6 Sol (Max) | 1682 |
| 9 | Gemini 4 Argon (High), Preliminary | 1678 |
| 10 | Qwen3.8-Max (0902), Preliminary | 1674 |
Open-weight standouts: DeepSeek-V4.1-Flash-20261017 (Max) #12 at 1670 (MIT), Kimi K3 (Max) #15 at 1655, MiMo-V2.6-Flash #18 at 1633 (MIT).
Sources: LMArena WebDev leaderboard
LMArena Vision
Data as of Oct 9, 2026, slightly older than the other two LMArena boards: 165 models.
| Rank | Model | Score |
|---|---|---|
| 1 | Claude Fable 5 (High) | 1308 |
| 2 | Qwen3.8-Max | 1301 |
| 3 | Claude Opus 4.6 (High) | 1299 |
| 4 | Claude Opus 4.7 | 1298 |
| 5 | Claude Opus 4.7 (High) | 1298 |
| 6 | Gemini 3.7 Flash (High), Preliminary | 1297 |
| 7 | Claude Opus 4.6 | 1294 |
| 8 | GPT-6.1 Sol (Max) | 1294 |
The top ten sit within about 15 points, and the printed intervals are ±7 to ±14, so most of these ranks overlap.
Sources: LMArena Vision leaderboard
Artificial Analysis Intelligence Index
Live page, no date shown, 134 models ranked. Artificial Analysis lists each reasoning-effort setting as its own row; we show the top rows as listed.
| Rank | Model (setting) | Index |
|---|---|---|
| 1 | Claude Opus 5.5 (max) | 58 |
| 2 | Claude Sonnet 5.5 (max) | 56 |
| 3 | Claude Opus 5.5 (xhigh) | 56 |
| 4 | Muse Wonder Full (max) (Meta) | 56 |
| 5 | Claude Opus 5.5 (high) | 54 |
| 6 | Claude Fable 5.1 (max) | 53 |
| 7 | Claude Fable 5.1 (xhigh) | 53 |
| 8 | GPT-6 Astra (max) | 53 |
| 9 | Gemini 4 Argon (high) | 53 |
| 10 | Muse Wonder (max) (Meta) | 52 |
Top open-weight models per the site's own FAQ: MiMo-V2.6-Pro 46, GLM-5.3 (Max) 45, Step 5 Preview 44. Other reference points on the same page: Kimi K3 (max) 44, Qwen3.8 Max (0902) 45 (the page doesn't say whether its weights are open), DeepSeek V4.1 Flash (max) 39. The best open model is 12 points behind the leader.
Sources: Artificial Analysis LLM leaderboard
LiveBench 2.0
Release 2026-10-14, last updated 18 October 2026. "Overall" is the mean of the suites a model has a score for.
| Rank | Model | Overall |
|---|---|---|
| 1 | Claude Fable 5.1 | 88.0 |
| 2 | Claude Opus 5.5 | 86.5 |
| 3 | GPT-6 Astra | 86.1 |
| 4 | GPT-6.1 Sol | 85.4 |
| 5 | Gemini 4 Argon | 84.1 |
| 6 | Claude Fable 5 | 83.6 |
| 7 | Claude Sonnet 5.5 | 82.5 |
| 8 | Claude Opus 5 | 82.1 |
| 9 | GPT-6 Sol | 81.9 |
| 10 | DeepSeek V4.1 Flash 1017 (marked open) | 79.2 |
Next open models: DeepSeek V4.1 Flash 78.0, MiMo V2.6 Pro 74.9, GLM-5.3 Flash 70.6. Category leaders include Gemini 4 Argon in Reasoning (95.1), Claude Fable 5.1 in Browser Use (100.0), Claude Fable 5 in Coding (98.3) and Claude Opus 5.5 in Agentic (85.0). Third-party sites sometimes still show the previous 2026-06-25 edition of LiveBench, so check the edition before comparing numbers.
Sources: LiveBench
Scale AI: Humanity's Last Exam (HLE)
Scale's HLE page, as fetched 19 Oct 2026. It carries a Sep 17, 2026 note about the HLE-Rolling fork. Scores are accuracy with the ± the page prints.
| Rank | Model | Accuracy (%) |
|---|---|---|
| 1 | GPT 6 Astra | 54.80±1.94 |
| 2 | Fable 5.1 (xhigh) | 46.50±2.00 |
| 3 | gemini-3.1-pro-preview (thinking high) | 46.44±1.96 |
| 4 | Sonnet 5.5 (xhigh) | 46.12±1.95 |
| 5 | Gemini 3.8 Flash | 44.52±1.96 |
| 6 | gpt-5.4-pro-2026-03-05 | 44.32±1.95 |
| 7 | Muse Spark | 40.56±1.92 |
The best open model on this page is kimi-k2.5 at 24.37, in 17th place. The page's methodology ranks by confidence interval, so close scores share a rank in effect. Note that Scale's main leaderboard page showed a different HLE card in our earlier read (GPT-6 Astra 60.60, Claude Opus 5.5 55.00, Gemini 4 Argon 53.20) and does not explain the difference, so we show only the detail page here. The HLE page does not list Opus 5.5 or Gemini 4 Argon.
Sources: Scale HLE leaderboard, Scale leaderboard hub
Scale AI: other boards (not re-verified in this pass)
From our earlier read on 19 Oct 2026. Harnesses differ by row (Claude Code, Codex, mini-SWE-agent), so don't compare across boards. Boards that show ranks without scores were left out.
| Board | Leaders (score) |
|---|---|
| Humanity's Sixth Sense | Gemini-4-Argon (high) 55.40, GPT-6-Astra (max) 53.60, GPT-6.1-Sol (max) 46.60 |
| SWE Atlas Refactoring | GPT 6 Astra (Codex) xHigh 59.05, Fable-5.1 (Claude Code) 56.67, Fable-5 54.76 |
| SWE Atlas Test Writing | Fable-5.1 (Claude Code) 67.04, Opus 5 (Claude Code) 62.22 |
| SWE Atlas Codebase QnA | Opus 5 (Claude Code) 63.17, Fable-5.1 59.95 |
| HiL-Bench | Muse Wonder 1.0 79.67, MuseSpark 1.3 66.89 |
| DrugDiscoveryBench | GPT 6 Astra (mini-SWE-agent) 68.70, GPT 6 Astra (Codex) 65.40, Muse-Spark 1.3 62.20 |
| Remote Labor Index | Opus 5.5 21.25 (1st) |
Sources: Scale leaderboard hub
Epoch AI: Epoch Capabilities Index (ECI)
We could not read the ECI table (it loads with JavaScript). The only figures we have come from Epoch's home-page text, announced Oct 1, 2026: Claude Opus 5.5 first at 167, GPT-6 Astra just behind, and Claude Sonnet 5.5 about level with Claude Fable 5.1 at 165. Epoch's HLE page showed a description but no scores, so we used Scale's HLE above. Epoch's ECI data file was labeled "Updated Oct. 18, 2026".
| Model | ECI (as stated on the home page) |
|---|---|
| Claude Opus 5.5 | 167 |
| GPT-6 Astra | slightly below Opus 5.5 (no number read) |
| Claude Sonnet 5.5 / Claude Fable 5.1 | about 165 |
Sources: Epoch AI benchmarks, Epoch Capabilities Index
SWE-bench Verified (third-party aggregators, not re-verified)
The official swebench.com page rendered no score table for us, so we could not check the official ranking. Aggregators show the following, and they mix vendor-reported with independently measured results. The benchmark is also close to saturated at the top.
| Rank | Model | Resolved (%) |
|---|---|---|
| 1 | Claude Fable 5 | 95.0 |
| 2 | Claude Mythos Preview | 93.9 |
| 3 | Ember-1 (Fireworks) | 92.2 |
| 4 | Claude Opus 4.8 | 88.6 |
| 5 | Claude Opus 4.7 | 87.6 |
| 6 | Claude Sonnet 5 | 85.2 |
(LLMBoard list, updated Oct 15, 2026. It continues down to 80.4 at rank 12.) A second aggregator, LLMLearner, listed Claude Opus 5 at 96.0 as a vendor-reported figure and Fable 5 at 95.0 as an independent measurement from vals.ai. The best open-weight entry there, Beam (Reflection AI) at 80.9, is the vendor's own number.
Sources: LLMBoard SWE-bench Verified, LLMLearner SWE-bench Verified rankings, official SWE-bench site
OpenRouter: usage, not quality
Usage data through Oct 17, 2026. These are tokens processed per week through OpenRouter's API. Cheap and free models are favored, and the site itself says the ranking is not a measure of accuracy or capability.
| Rank | Model | Author | Weekly tokens | Change |
|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | DeepSeek | 35.9T | -7% |
| 2 | Step 5 Preview | StepFun | 13.9T | +119% |
| 3 | GLM 5.3 Flash | Z.ai | 12.6T | +8% |
| 4 | MiMo-V2.6-Flash | Xiaomi | 10.3T | -9% |
| 5 | Hy4 preview | Tencent | 7.55T | -10% |
| 6 | GPT-6 Luna | OpenAI | 6.29T | -12% |
| 7 | Claude Opus 5.5 | Anthropic | 6.05T | +29% |
| 8 | Nemotron 3 Ultra (free) | NVIDIA | 5.34T | -19% |
| 9 | GLM 5.3 | Z.ai | 5.34T | +55% |
| 10 | DeepSeek V4 Flash 0731 | DeepSeek | 4.71T | -7% |
Monthly, DeepSeek V4.1 Flash leads with 125T tokens, followed by a stealth model, Space Bunny Alpha (56.7T), and GLM 5.3 Flash (56.2T). By share of requests in the week beginning Oct 5: DeepSeek 24.5%, OpenAI 18.6%, Google 17.8%, Z.ai 13.4%, Qwen 5.4%, Anthropic 3.9%.
Sources: OpenRouter rankings
Stale boards: don't use these for current rankings
Aider Polyglot (last updated November 20, 2025)
About eleven months old, with none of the 2026 models. Shown for reference only.
| Model | Score |
|---|---|
| gpt-5 (high) | 88.0% |
| gpt-5 (medium) | 86.7% |
| o3-pro | 84.9% |
| gemini-2.5-pro-preview-06-05 (32k think) | 83.1% |
| gpt-5 (low) | 81.3% |
| o3 (high) | 81.3% |
| grok-4 | 79.6% |
| DeepSeek-V3.2-Exp Reasoner | 74.2% |
| Kimi K2 | 59.1% |
Sources: Aider LLM leaderboards
Scale SWE-Bench Pro, public set (no date; models stop around March 2026)
| Rank | Model | Score (%) |
|---|---|---|
| 1 | Muse Spark 1.1 | 61.50±3.10 |
| 2 | gpt-5.4 (xHigh) | 59.10 |
| 3 | Muse Spark | 55.00 |
| 4 | claude-opus-4-6 (thinking) | 51.90 |
| 5 | gemini-3.1-pro | 46.10 |
| 6 | claude-opus-4-5 | 45.89 |
| 7 | claude-4-5-Sonnet | 43.60 |
Opus 5, Fable and GPT-6 models are not on this page, so it says nothing about the current frontier.
Sources: Scale SWE-Bench Pro (public)
Takeaways
- Leaders differ by board. Gemini 4 Argon is first on LMArena text, Anthropic's Opus 5.5 / Fable 5.1 lead Artificial Analysis, LiveBench and WebDev, GPT-6 Astra leads HLE, and Fable 5 leads Vision. Anthropic models fill most of the top rows of most boards, while OpenAI and Google win individual boards.
- LMArena ranks are softer than they look. Gemini 4 Argon's first place is marked Preliminary, its ±9 interval overlaps those of #2 and #3, and the site's rank spread for it is 1 to 3. Gaps of a few points between neighbors are effectively ties. We also could not tell from the page whether style control is applied.
- Vendor-reported is not independent. Artificial Analysis, LMArena, LiveBench and Scale run their own evaluations. SWE-bench aggregators blend vendor system-card numbers with independent runs, and Scale's coding boards use different agent harnesses, so cross-board comparison is shaky.
- OpenRouter measures usage, not quality. Its top spots go to cheap, fast models, and the site says so itself.
- Open and Chinese-lab models. At the very top, proprietary models dominate: on Artificial Analysis the best open model (MiMo-V2.6-Pro, 46) trails the leader (58) by 12 points, and the top 40 of LMArena text is almost all proprietary. They do better on specific boards: DeepSeek V4.1 Flash is 12th on WebDev and 10th on LiveBench, Kimi K3 is 18th on LMArena text, and DeepSeek, GLM, MiMo and Step models dominate OpenRouter usage.
What we couldn't verify
- Official SWE-bench and the Epoch ECI table could not be read (JavaScript-rendered), and Epoch's HLE page had no scores.
- Stale: Aider Polyglot (Nov 2025) and Scale SWE-Bench Pro (about March 2026).
- Not re-fetched in this pass: the Scale "other boards" table, the SWE-bench aggregator numbers and the Aider and SWE-Bench Pro tables come from our earlier read of the same day.
- Not checked at all: METR, Design Arena, Terminal-Bench and Vals.ai directly.
- Leaderboards change often. Treat everything here as a snapshot from 19 Oct 2026, KST.
← Back to all posts