Best AI Model Right Now (November 2026): LMArena, Artificial Analysis and WebDev Leaderboards Explained

“Which AI is best right now?” has a different answer depending on which leaderboard you read and what you want the model to do. In early November 2026 Google’s Gemini 4 Carbon tops the LMArena text chart, Anthropic’s Claude Opus 5.5 leads coding and the composite indexes, and OpenAI’s GPT-6 Astra sits close behind in both. Open-weights models such as Kimi K3, GLM-5.3 and DeepSeek V4 Pro are closer to the top than ever. Below I pull the latest numbers from the main leaderboards into one place, explain what each one actually measures, and give a plain recommendation by task.

LMArena Text Leaderboard (Top 20)

LMArena ranks models by blind head-to-head votes from real users. Data via BenchLeader as of November 5, 2026 (results published October 28):

Two things stand out. First, the top 20 is separated by only about 40 points, which in Arena terms means the top models win only slightly more than half their head-to-head votes against each other. Second, Gemini 3.8 Flash at #10 is remarkable for a fast, cheap model. GPT-6 models do not appear in this text top 20 snapshot, though they rank highly on other leaderboards below. Open-weights entries in the top 30: Kimi K3 (#19), GLM 5.3 (#28) and Xiaomi MiMo-V2.6-Pro (#29).

More on the leader in my Gemini 4 Carbon benchmarks and access post.

LMArena WebDev (Coding) Leaderboard

WebDev Arena has users compare models building real web apps. As of November 4, 2026:

Claude Opus 5.5 has a clear lead of about 37 points here, and Sonnet 5.5 at #3 is the value pick for coding. If you code with agents, see my Claude Code cheat sheet and OpenAI Codex guide.

Artificial Analysis Intelligence Index

Artificial Analysis combines many benchmarks (reasoning, knowledge, math, coding) into one score. Top five as of November 2, 2026, via ArtificialWatch:

On ArtificialWatch’s own composite, Claude Opus 5.5 ranks first (98.1), ahead of Gemini 4 Carbon (94.8) and GPT-6 Astra (92.6). In other words: among models you can actually use today, Claude Opus 5.5 is the top all-rounder by both measures.

SWE-bench Coding Scores

SWE-bench tests whether a model can fix real GitHub issues. Reported scores collected by Swfte:

Verified scores are now crowded above 80%, which is why many labs report the harder SWE-bench Pro instead. Many of the newest frontier models do not have consistent public SWE-bench numbers in this tracker yet, so use WebDev Arena for a fresher comparison.

Best Open-Weights Models

Most of these are too big for a single consumer GPU, but smaller siblings and quantized versions run locally. See my guides to running Kimi locally, GLM Flash locally and the best Ollama models.

What Each Leaderboard Measures

• LMArena (text): human preference in blind chats. Great for “which answer do people like”, but it rewards style and length, and the gaps at the top are small.
• LMArena WebDev: human votes on working web apps. A good proxy for practical frontend coding.
• Artificial Analysis Index: automated benchmarks combined. More objective, but benchmarks can be trained toward.
• SWE-bench: real bug fixes in Python repos. Verified is near saturation; Pro is harder.
• Reasoning settings matter: labels like (high), (max) or (xhigh) mean the score used extra thinking time, which costs more and is slower than the default.

Benchmarks also age fast. A November 2026 Stanford study found 44% of widely cited LLM findings had expired within about 16 months; see my write-up.

Best Model by Task

Best Value: Score per Dollar

Top scores come at top prices. GPT-6.1 Sol at $2 input and $10 output per million tokens scores 54.5 on the AA index, close to GPT-6 Astra’s 56.9 at $10 and $50. Gemini 3.8 Flash ranks #10 on LMArena text at Flash pricing. Full price tables: my AI API pricing comparison, Claude API pricing, OpenAI API pricing and Gemini API pricing.

How to Pick the Right Model for Your Task in 5 Steps

Leaderboards are a starting point, not a verdict. This is the process I’d use to turn the tables above into a decision for a real project:

1. Write down the exact task. “Coding” is too vague. “Fix failing tests in a large TypeScript repo” or “summarise 40-page contracts” points you to very different models and benchmarks.
2. Pick the leaderboard that matches it. Use WebDev Arena for frontend and app building, SWE-bench Pro for bug fixing in real repositories, the Artificial Analysis index for broad reasoning and knowledge, and LMArena text for general chat quality.
3. Shortlist three models, including one cheap one. For example Claude Opus 5.5, GPT-6.1 Sol and Gemini 3.8 Flash. Only consider models you can actually call today, which rules out Gemini 4 Carbon for now.
4. Run 20 to 30 of your own prompts. Use real examples from your work, hide which model wrote each answer, and score the results against a simple checklist. Your own test beats any public score for your use case.
5. Compare cost per finished task. Multiply tokens used by price, and add the cost of retries and fixes. A cheaper model that needs two attempts can cost more than a premium model that gets it right first time.

Repeat the test when a major new model ships. With rankings changing weekly, a quick re-run every couple of months keeps you from overpaying.

Common Problems When Comparing Models and Fixes

The leaderboard winner does worse on my task

Each leaderboard measures something narrow. A model that wins blind chat votes may not be the best at strict JSON output, long documents or your programming language. Fix: treat the leaderboard as a shortlist and decide with your own prompts, as in the steps above.

Arena votes reward style

Human voters tend to prefer longer, well-formatted answers, even when a shorter one is more correct. LMArena has published style-controlled rankings that adjust for this. Fix: if you care about accuracy over presentation, weight the automated benchmarks more heavily and check facts in the answers yourself.

Benchmark contamination

Public test questions can leak into training data, so a high score may partly reflect memorisation. This is one reason labs moved from SWE-bench Verified, now crowded above 80%, to the harder SWE-bench Pro. Fix: prefer newer or private benchmarks and your own unseen examples.

Score gaps that are too small to matter

The LMArena text top 20 spans only about 40 points, so a few points between neighbours is close to noise. Fix: treat models within a handful of points as tied and choose on price, speed and features.

Price versus score confusion

Scores with labels like (high), (max) or (xhigh) use extra reasoning, which means more output tokens and higher bills than the default setting. Fix: check which setting produced the score, then price that setting, not the base model.

The API model name differs from the leaderboard name

Leaderboards combine a model name with a reasoning setting, such as “Claude Opus 5.5 (high)”, while APIs use model IDs such as claude-opus-5-5 and set reasoning effort as a separate parameter. Fix: look up the exact model ID in the provider’s docs and match the effort setting to the one the leaderboard used.

The top model is not available yet

Gemini 4 Carbon leads LMArena text and the AA index but is not generally available. Fix: compare among models you can use today, where Claude Opus 5.5 and Gemini 4 Argon lead.

Frequently Asked Questions

What is the best AI model in November 2026?

Among generally available models, Claude Opus 5.5 leads coding and the composite indexes, while Gemini 4 Argon is #2 on LMArena text. Gemini 4 Carbon tops LMArena text but is not generally available yet.

Is ChatGPT or Claude better right now?

On these leaderboards, Claude Opus 5.5 edges GPT-6 Astra in coding (WebDev 1810 vs 1773) and on the AA index (58.4 vs 56.9). For a full product comparison, see my ChatGPT vs Claude vs Gemini vs Grok guide.

What is the best open-source AI model?

Kimi K3 ranks highest among open-weights models on LMArena text, with GLM-5.3 and DeepSeek V4 Pro close behind.

How often do these leaderboards change?

Weekly. New models and reasoning modes shift positions all the time, so check the date on any ranking.

Why do LMArena and Artificial Analysis rank models differently?

They measure different things. LMArena counts human preference votes in blind chats, while the Artificial Analysis index combines automated benchmarks for reasoning, knowledge, math and coding. That is why Gemini 4 Argon ranks #2 on LMArena text while Claude Opus 5.5 leads the AA index among available models.

Should I trust SWE-bench Verified scores?

As one signal, yes, but top models are now bunched above 80%, so Verified no longer separates them well. Look at SWE-bench Pro and WebDev Arena too, and test on your own codebase before committing.

Sources

• BenchLeader: LMArena Text leaderboard (November 5, 2026)
• BenchLeader: LMArena WebDev leaderboard (November 4, 2026)
• ArtificialWatch: LLM leaderboard with Artificial Analysis data (November 2 to 4, 2026)
• Swfte: AI model leaderboard November 2026
• AF: Gemini 4 Carbon leads LMArena text with 1528