A 10-minute tour · data read November 5, 2026
Who's ahead in AI, what it costs, and who actually uses it
Six short sections on the state of large language models in November 2026, each with one question to discuss. Every score on this page comes from one composite index, so the numbers can be compared with each other.
60
Top score: Gemini 4 Carbon
But cyber defenders only. Best model you can buy: Claude Opus 5.5 at 59.
13 pts
Closed lead over open
Claude Opus 5.5 59 vs Kimi K3.1 Pro 46, the best open-weight model.
11×
Cost spread in the top 10
$0.96 per task (GPT-6.1 Terra) to $10.85 (Claude Fable 5.1).
72%
Open-weight share of tokens
On OpenRouter, week of Sep 28, 2026. Total volume grew 25× in a year.
1 · Leaderboard
The top is crowded, and the top model isn't for sale
Artificial Analysis Intelligence Index v5, each model at its best reasoning setting
A composite index averages many tests into one number. This one combines nine evaluations across knowledge work, coding, science and general knowledge, and 60% of it uses private test sets that labs can't train on.
Five points separate the top five. Anthropic, OpenAI and Google hold every spot at 50 or above except Meta's Muse Wonder Full — which, like the leader, you can't call through an API.
The best open-weight model, Kimi K3.1 Pro, sits 13 points behind the best model you can buy. Its weights went public on November 4.
Discuss. Gemini 4 Carbon is first, but only a few security teams can use it. Should a leaderboard rank models the public can't buy? What would you put on a ranking you showed your CEO?
2 · Cost per task
Score is half the story — the other half is what one answer costs
Index score vs cost per index task. Dashed line: best value at each budget. Tap a dot for its name.
Cost per task is what the model actually spent to complete one test item: the tokens it read and wrote, including its hidden reasoning, times its list prices.
The dashed line marks the models nobody beats on score for less money, leaving out models you can't buy: GPT-6 Luna, MiMo-V2.6-Pro, Claude Haiku 5.5, GPT-6.1 Terra, GPT-6.1 Sol, GPT-6 Astra, Claude Opus 5.5. Everything below the line is dominated — something else scores higher for less.
GPT-6.1 Sol scores 55 for $1.27. Claude Opus 5.5 scores 59 for $10.04. Four extra points cost about eight times as much — worth it for some work, not for most.
Discuss. You're building a customer-support assistant that will handle a million tickets a month. Which dot would you pick, and what would have to be true to justify moving up and to the right?
3 · Tokens vs tasks
A cheap token is not a cheap answer
- Claude Haiku 5.5score 36Per token$0.50/1MPer task$0.41
- DeepSeek V4.1 Flashscore 33Per token$1.20/1MPer task$0.60
- Gemini 3.8 Flashscore 31Per token$3.75/1MPer task$7.77
- GLM-5.3score 32Per token$4.40/1MPer task$5.01
- GPT-6.1 Terrascore 51Per token$5.00/1MPer task$0.96
- Grok 4.7score 35Per token$6.00/1MPer task$6.18
- Qwen3.8-Maxscore 34Per token$6.00/1MPer task$10.02
- GPT-6.1 Solscore 55Per token$10.00/1MPer task$1.27
| Model | List price, output ($ per 1M tokens) | Actual cost per task | Index score |
|---|---|---|---|
| Claude Haiku 5.5 | $0.50 | $0.41 | 36 |
| DeepSeek V4.1 Flash | $1.20 | $0.60 | 33 |
| Gemini 3.8 Flash | $3.75 | $7.77 | 31 |
| GLM-5.3 | $4.40 | $5.01 | 32 |
| GPT-6.1 Terra | $5.00 | $0.96 | 51 |
| Grok 4.7 | $6.00 | $6.18 | 35 |
| Qwen3.8-Max | $6.00 | $10.02 | 34 |
| GPT-6.1 Sol | $10.00 | $1.27 | 55 |
Discuss. Gemini 3.8 Flash charges under 40% of GPT-6.1 Sol's token price and costs six times as much per task. If you were a vendor, would you price per token, per task, or per outcome — and what would each do to your customers' behavior?
4 · The effort dial
The same model gets smarter if you pay it to think longer
Each line is one model at five reasoning settings, low → max. Same index, same cost measure as above.
Modern models reason before answering, and you can set how long. More reasoning means more tokens, which means a higher bill.
Claude Opus 5.5 goes from 35 to 59 as effort rises — and from $0.98 to $10.04 per task. GPT-6.1 Sol tops out lower, at 55, but gets there for $1.27. Opus scores 52 at $3.46 and needs $6.05 (xhigh, 57) to beat it.
A model's "score" is really a curve. Two companies using the same model can get very different quality and cost depending on this one setting.
Discuss. Effort is a setting you choose on every request. How would a company decide which tasks get 'max' and which get 'low'? Who should own that decision — engineering, finance, or the person asking?
5 · Open vs closed
Open models trail on capability and lead on usage
Share of weekly tokens on OpenRouter, top 50 models
Open weights means anyone can download the model and run it, so many companies compete to serve it. That competition drives price down.
On OpenRouter, open models averaged $1.62 per million output tokens against $6.36 for closed — about four times cheaper. Their share of tokens went from 27% to 72% in a year, while total volume grew about 25×.
Closed volume still grew almost tenfold. Tokens are not revenue: cheap models take the bulk work, and premium models keep the tasks where a wrong answer is expensive.
Discuss. Open-weight models are 13 points behind but handle about 72% of OpenRouter's tokens. Who is choosing them, for what work — and does the capability gap matter for that work?
6 · Reading a score
What's inside the number — and why it keeps changing
The index version
Xiaomi's MiMo-V2.6-Pro launched in September as the top open model at 46 on the previous version of this index. On v5 it reads 30. The model didn't change; the test did.
The benchmark version
Gemini 3.8 Flash scored 87.6% on Terminal-Bench 2.1 and 19.1% on 4.0. Popular tests fill up, labs tune for them, and the field moves to harder ones.
Who ran it
Launch posts report the lab's own numbers, on settings the lab picked. This page uses one independent tester for every score. The benchmark guide covers the rest.
Discuss. Labs know which tests the market watches. If you were buying, would you trust a public benchmark, a private one, or your own test on your own data? What would your own test look like?
Next
From chatbots to always-on agents
Between May and September, four big labs launched agents that run on their own cloud computer and keep working after you close the app: OpenAI Dots, Meta Muse, Google Gemini Spark and SpaceXAI's Grok Bot. What they do, what they cost, and how they ask permission.