AI Leaderboards

0:00 There is no single "best AI model" ranking, but the leaderboards agree more than they differ. On Nov 6, 2026, Google's Gemini 4 Carbon was first on LMArena's Text leaderboard (1528, preliminary, as of Oct 27, 2026) and on the Artificial Analysis Intelligence Index (60), with Anthropic's Claude Opus 5.5 one point behind on the index (59) and first on LMArena WebDev for coding. OpenAI's GPT-6 Astra placed second on WebDev and fifth on the index. Among models whose weights you can download, Moonshot's Kimi K3.1 Pro scored highest on the index (46).

3:56 Use the leaderboards to shortlist, not to decide. Each measures something different: LMArena measures which answers people prefer, Artificial Analysis averages hard tests, and SWE-bench and Terminal-Bench measure coding agents. A model that tops one can trail on your task, and makers sometimes tune models for a leaderboard. Try two or three on your own work. The task-by-task picks are in Best AI Chatbot, Best AI Coding Agent and Best AI Image Generator.

The leaderboards at a glance

All read Nov 6, 2026. LMArena scores dated Oct 27, 2026, its last update.

LMArena (Arena)

2:38 LMArena, now at arena.ai under the name Arena, shows people two anonymous answers to their own prompt and asks which is better; the votes become an Elo-style score with a margin of error. It covers text, web development, vision, image generation and editing, video, search and documents. Its site counted 374,859,400 sessions on Nov 6, 2026. In a Feb 10, 2026 video, Arena showed its image, search and video arenas working the same way: one prompt, two anonymous results, and a vote for the better one.

Text, top 10 (Oct 27, 2026):

Source: LMArena Text leaderboard, read Nov 6, 2026. Scores within a few points overlap in their margins of error, so places 4 to 9 are close to a tie.

1:02 WebDev, top 5 (Oct 27, 2026): Claude Opus 5.5 (Max) 1810, GPT-6 Astra (Max) 1773, Claude Sonnet 5.5 (xHigh) 1770, Gemini 4 Carbon (High) 1768 (preliminary), GPT-6.1 Sol (Max) 1751 (LMArena WebDev). Voters there compare two working apps built from the same prompt, as Arena's own Feb 10, 2026 video shows.

Limits. Votes reward answers people like, which can favor length, formatting and a confident tone over accuracy. The clearest case: Meta's Llama 4 Maverick reached a high place in Apr 2025 with an unreleased chat-tuned version, and TechCrunch reported that the released model ranked well below it after LMArena changed its rules. See Llama.

Artificial Analysis

Artificial Analysis, an independent company, runs the same tests on every model itself and combines them into the Intelligence Index, now version 5.0.1: four equally weighted parts, Knowledge Work (AA-Briefcase, AutomationBench, GDP.pdf), Coding (FrontierCode 2.0, Terminal-Bench 4.0), Science (Terminal-Bench-Science, CritPt) and General (AA-Omniscience, HLE-Diamond). It also measures price and speed for each model and provider, which no vote-based board does.

Intelligence Index, top 10 (read Nov 6, 2026):

Source: Artificial Analysis model leaderboard. The words in brackets are reasoning effort: the same model thinking longer scores higher and costs more. Among models whose weights are published, Kimi K3.1 Pro (on Hugging Face) scored 46, DeepSeek V4.1 Flash 1017 (max) 33 and GLM-5.3 (max) 32 (MIT licensed), well behind the closed leaders; see Kimi, DeepSeek and GLM.

Limits. An average hides strengths: a model can lead the index and trail on the one skill you need. Read the category scores.

Coding leaderboards: SWE-bench and Terminal-Bench

SWE-bench gives a coding agent a real GitHub issue from a Python project and checks whether its fix passes the project's tests; its site has Verified (a human-checked subset), Multilingual, Multimodal, Lite and Full boards, and a Bash Only view that compares models with the same simple agent. Terminal-Bench, now version 4.0, measures agents completing tasks in a real terminal: building software, setting up servers, fixing systems. Both leaderboards' tables load in the browser, and neither returned rows when we read them on Nov 6, 2026, so we do not quote a leader; Terminal-Bench 4.0 is also part of the Artificial Analysis index above. Makers cite these tests in launch posts (for example Mistral's 28.3% on Terminal-Bench 4 for Mistral Large 4), but a maker's own figure uses its own agent setup and is not the same as an independent run.

Limits. Agent scores depend on the agent around the model (tools, time and attempts allowed), so the same model can post very different numbers. Compare entries that use the same agent.

Humanity's Last Exam

Humanity's Last Exam (HLE), from the Center for AI Safety and Scale AI, is 2,500 hard questions written by nearly 1,000 experts across more than 100 subjects, finalized in Apr 2025, made to be harder than older tests that models had saturated. A harder subset, HLE-Diamond, followed on Sep 22, 2026. On Artificial Analysis's own run, Claude Opus 5.5 (Max) scored 61.4%, ahead of Claude Fable 5.1 at 59.1% (read Nov 6, 2026). HLE's own site still shows older results (Gemini 3 Pro at 38.3%), so use a dated, current run.

Limits. A written exam measures knowledge and reasoning, not whether a model is pleasant or useful for everyday tasks.

Usage rankings are not quality rankings

OpenRouter's rankings count tokens sent through its API: through Nov 5, 2026, DeepSeek V4.1 Flash 1017 led with 56.5 trillion tokens, and DeepSeek had 29.8% of requests in the week from Oct 26. That shows price and popularity with developers' coding agents, not ability. See OpenRouter.

How to read any leaderboard

• Check the date. A ranking from three months ago is history in this field; every number on this page carries the date read.
• Check the margin. Two LMArena scores within each other's error bars are a tie.
• Check the setting. Reasoning effort, the agent, and the number of attempts change scores a lot.
• Prefer independent runs. A maker's model card is useful but is the maker's own measurement.
• Test on your task. Ten of your own prompts beat any average.

How we chose

This page uses leaderboards that publish their method and date their results: LMArena (human preference), Artificial Analysis (independent test runs, price and speed), SWE-bench and Terminal-Bench (coding agents), Humanity's Last Exam (expert questions) and OpenRouter (usage). Each was read on Nov 6, 2026; where a table would not load, we say so instead of quoting a stale figure. Makers' own benchmark claims are labeled as theirs.

Frequently asked questions

What is the best AI model right now? As of Nov 6, 2026, Google's Gemini 4 Carbon led both LMArena Text and the Artificial Analysis Intelligence Index, with Anthropic's Claude Opus 5.5 a point behind on the index and first for web coding on LMArena WebDev.

What is LMArena? LMArena, now at arena.ai, is a public leaderboard where people compare two anonymous AI answers and vote for the better one; the votes produce an Elo-style ranking for text, coding, images, video and more.

What is the Artificial Analysis Intelligence Index? The Artificial Analysis Intelligence Index (v5.0.1) averages nine tests in knowledge work, coding, science and general knowledge, run by Artificial Analysis itself; Gemini 4 Carbon led it with 60 on Nov 6, 2026.

What is the best open-weight model? On the Artificial Analysis Intelligence Index read Nov 6, 2026, Moonshot's Kimi K3.1 Pro scored highest among open-weight models we could confirm, with 46, ahead of DeepSeek V4.1 Flash 1017 (33) and GLM-5.3 (32).

Which AI is best for coding? Claude Opus 5.5 led LMArena's WebDev leaderboard (1810, Oct 27, 2026) and scored second on the Artificial Analysis Intelligence Index; see Best AI Coding Agent for tools.

Can AI leaderboards be gamed? Yes. Makers can tune models for a test or submit special versions, as happened with Llama 4 Maverick on LMArena in Apr 2025; use several leaderboards and your own tests.

Sources

• Text leaderboard, LMArena, read Nov 6, 2026
• WebDev leaderboard, LMArena, read Nov 6, 2026
• Leaderboard, LMArena, read Nov 6, 2026
• Arena leaderboards explained: chat, code, vision & more, Arena, Feb 10, 2026 (video)
• Model leaderboard, Artificial Analysis, read Nov 6, 2026
• Intelligence benchmarking methodology, Artificial Analysis, read Nov 6, 2026
• Humanity's Last Exam, Artificial Analysis, read Nov 6, 2026
• Humanity's Last Exam, Center for AI Safety and Scale AI, read Nov 6, 2026
• Terminal-Bench 4.0 leaderboard, Terminal-Bench, read Nov 6, 2026
• SWE-bench, read Nov 6, 2026
• Rankings, OpenRouter, read Nov 6, 2026
• Introducing Mistral Large 4, Mistral AI, Oct 6, 2026
• Meta's vanilla Maverick AI model ranks below rivals on a popular chat benchmark, TechCrunch, Apr 11, 2025