Artificial Analysis is an independent benchmarking company, and its LLM leaderboard ranks more than 250 models on intelligence, cost, output speed, latency and context window. As of November 4, 2026, its Intelligence Index puts Gemini 4 Carbon at high effort on top with 60, one point ahead of Claude Opus 5.5 at max effort.
That single number is what most people quote. For anyone who pays per token, the columns next to it matter more.
What each column on the board means #
| Column | What it measures | How to use it |
|---|---|---|
| Intelligence Index | Weighted average of nine evaluations in v5.0.1, a quarter each for knowledge work, coding, science and general | A rough rank of overall ability, not of your task |
| Cost per Task | Weighted average dollars spent per index task | The closest thing to a real bill for hard work |
| Output speed | Median tokens per second | How fast text streams once it starts |
| Latency, first chunk | Seconds before the first output arrives | How long a user stares at a spinner |
| Total response | Seconds for the whole answer | Includes thinking time |
| Context window | Largest input the model accepts | How much code or text you can send at once |
Every model is tested under the same prompts, temperature and grading, according to the methodology page, so rows compare fairly with each other. They do not tell you how a model handles your repository or your customers.
One model, several rows #
The board lists the same model at different effort settings, and the spread is wide. Claude Opus 5.5 scores 59 at max effort for $10.04 a task, and 52 at high effort for $3.46. Gemini 4 Carbon scores 60 at high effort and 58 at medium, and the wait changes more than the score: about 152 seconds to the first chunk at high against about 18 at medium.
Then there is the value row. GPT-6.1 Sol at max effort scores 55 for $1.27 a task, against $7.76 for Carbon's top setting. Giving up five points of index for about a sixth of the price is a trade many teams will take for routine work. Read effort labels before you read scores.
Other boards on the site #
The leaderboard is one page of several. A Coding Agent Index combines FrontierCode, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA; in its October 28 write-up, Artificial Analysis put Carbon second there, behind Opus 5.5. A separate open source page lists models whose weights you can download, led on November 4 by DeepSeek V4.1 Flash 1017 and GLM-5.3 at max effort. There is also a Model Recommender that weighs intelligence, speed and cost, and a tool called Optima for building a custom benchmark.
The index itself moves. Version 5 launched on October 27, Carbon took first place the next day, and a regrade on November 3 produced v5.0.1. A score from last month may not match today's board.
Where TechTwitter IO fits #
TechTwitter IO, a free satirical daily paper about tech Twitter, opens its own Benchmarks page by naming Artificial Analysis as the place that measures how smart a model is, then ranks what nobody measures: launch claims against outside proof, "I switched" posts, vibes, teaser hype and model misdeeds. Its Trust Me Bro board, for example, adds points for launch-chart fine print and for charts whose y-axis does not start at zero, then divides by outside proof.
It is the pick for a weekly laugh at the claims around each launch, with every score linked to the posts behind it. It is also satire by its own footer, so its numbers never belong in a model decision. Use Artificial Analysis for the measurements and the paper for the commentary.
When to look elsewhere #
If you care which answer people prefer rather than which model scores higher on tests, Arena's blind-vote boards answer that better. And no public index replaces twenty of your own prompts run against two or three finalists.