AI Model Rankings November 2026: Who Actually Leads

AI leaderboards are useful to buyers. Chasing a single #1 number without knowing what it measures is how AI budgets get wasted. In November 2026 the honest answer to "who leads?" is "leads at what?" Anthropic tops the composite tables, Google tops the human-preference arena, and OpenAI is close behind on nearly every board. The model you actually need depends on your product, not on the headline.

How to read a leaderboard

Three kinds of leaderboard dominate the conversation, and they answer three different questions. Treating them as the same thing is the first mistake most teams make.

Composite indexes, like the LLMBenchmark.ai Index, combine many public benchmarks (coding, agentic work, reasoning, maths, knowledge) and outside scores into one number. They are designed to answer "which model is strongest across the board?" They are also the most opinionated. Change the weighting and the order changes.

Intelligence indexes, like the Artificial Analysis Intelligence Index, run a fixed set of hard evaluations under controlled conditions, often at maximum reasoning effort. They measure raw capability at full effort, not how the model behaves at the effort level you can afford in production.

Human-preference arenas, like LMArena's Text Arena, ask real people to pick the better of two anonymous answers and turn the votes into Elo ratings. They tell you what people like reading. That matters for consumer-facing copy and chat, and much less for a pipeline that has to call an API correctly ten thousand times a day.

Each one is a valid answer to its own question. None of them on its own tells you which model to buy.

The November 2026 rankings

Here is the top ten on the LLMBenchmark.ai composite index as read on 2 November 2026, with the Artificial Analysis Intelligence Index alongside where it has been published (high or max effort, as mirrored on LLMBenchmark.ai). A dash means that board has not published a score. We have not estimated any.

Rank	Model	Provider	LLMBenchmark Index	AA Intelligence
1	Claude Opus 5.5	Anthropic	86.9	58.4
2	GPT-6 Astra	OpenAI	83.9	56.9
3	Claude Fable 5.1	Anthropic	83.5	49.7
4	Claude Opus 5	Anthropic	82.7	—
5	Gemini 4 Argon	Google	82.6	50.4
6	GPT-6.1 Sol	OpenAI	82.0	54.5
7	Claude Sonnet 5.5	Anthropic	81.7	54.2
8	Claude Fable 5	Anthropic	78.9	—
9	GPT-6.1 Terra	OpenAI	76.8	51.0
10	GPT-6 Sol	OpenAI	76.4	—

Two things stand out. First, only three labs make the top ten: Anthropic, OpenAI and Google. Second, the gap is small. Ranks two to seven sit within about two points of each other, and the published error margins on those scores are bigger than that. Rank five and rank six are a statistical tie, not a verdict.

The other boards shuffle the order. On the Artificial Analysis Intelligence Index, Gemini 4 Carbon (high) leads with 59, just ahead of Claude Opus 5.5 (max) at 58 and GPT-6 Astra (max) at 57. On LMArena's Text Arena in late October, Gemini 4 Carbon (High) led on human preference at around 1528 Elo, followed by Gemini 4 Argon (High) and Claude Opus 5.5 (High).

By provider flagship

Anthropic: Claude Opus 5.5. The composite leader at 86.9, and first on LLMBenchmark's coding index. Anthropic has five of the top ten, and Claude Fable 5.1 leads the agentic board. If your work is code, agents or long multi-step tasks, Anthropic is the default to beat.

OpenAI: GPT-6 Astra. Second overall at 83.9, first on knowledge, and consistent across every category. The catch is price: about $50 per million output tokens, against $20 for Opus 5.5. GPT-6.1 Terra, at around $5, is the better-value OpenAI model in the top ten.

Google: Gemini 4 Carbon and Gemini 4 Argon. Carbon leads both the Artificial Analysis index and the human-preference arena. Argon scores 82.6 on the composite, leads reasoning, and costs about $10 per million output tokens. On quality per pound, Google is the strongest of the three big labs right now.

xAI: Grok 4.7. A clear tier below the frontier. Grok 4.7 sits around 63.5 on the composite and 35 on the Artificial Analysis index, and its predecessor Grok 4.6 scores higher on the composite at about 67.0. Grok has its uses, such as very long context and strong knowledge scores, but it is not a top-tier pick this month.

The open and efficient tier. This is where most production traffic should run. DeepSeek V4.1 Flash is the best open-weights model at 70.2, for about $0.90 per million output tokens. Claude Haiku 5.5 scores 64.2 on the composite (36 on Artificial Analysis at max effort) at around $0.50 and ships answers at roughly 238 tokens per second. OpenAI's GPT-5.6 Luna solves 93% of SWE-bench Verified tasks for about four cents per solved task. Kimi K3.1 Pro (74.4), Meta's Muse Spark 1.1 (73.3) and Qwen3.8 Max (67.6) round out a tier that would have been frontier-class eighteen months ago.

What the numbers miss

Benchmarks measure models. Your customers experience products. Between the two sits everything that decides whether AI makes you money.

Latency. A model that thinks for twenty seconds wins on maths and loses at checkout. Users do not wait.

Cost at volume. Price per token is not price per outcome. A cheap model that fails costs you the retry, the support ticket and sometimes the customer. Measure cost per successful task.

Tool use in production. Agentic benchmarks are getting better, but your CRM, your payment stack and your old internal APIs are not on any of them. Reliability on your own integrations is the score that matters.

Brand voice. Arena Elo rewards answers that read well to strangers. Your brand needs answers that sound like you, every time, at scale. That comes from prompting, evaluation and editorial control, not from the leaderboard.

Data residency and compliance. For UK and EU businesses, and especially anyone handling payments, where data is processed and retained can rule a model out before quality is even considered.

Team skill. A well-instrumented mid-tier model in the hands of a team that knows how to evaluate it will beat a frontier model bolted on by a team that doesn't. Every time.

How to pick for your stack

1. Route, don't marry. Use a frontier model such as Opus 5.5, GPT-6 Astra or Gemini 4 Carbon for the hard tenth of your requests, and an efficient model such as Haiku, DeepSeek Flash or Luna for the rest. One-model stacks are how you overpay.
2. Build your own eval before you sign anything. Fifty real tasks from your business, scored blind, will tell you more than every public leaderboard combined.
3. Ignore gaps of less than three points. Near the top, ranking differences are mostly noise. Choose on price, latency and integration fit instead.
4. Match the board to the job. Coding and agents: check the composite and agentic indexes. Customer-facing copy: check the arena. Hard analysis: check the intelligence index.
5. Re-test every quarter. This month's leader can be next quarter's third place. Design your stack so switching models is a config change, not a rebuild.

Pick a partner, not a logo

Choosing a model takes an afternoon. Making it pay off takes product design, marketing systems and engineering that holds up in production. At Eptune we have been building growth systems and payment infrastructure since 2011 for companies that need results, not demos. If you want AI wired properly into your product and your marketing, measured against your own numbers, talk to us.