Overall leaderboard
Full ranking →Ranked by the Artificial Analysis Intelligence Index v5 — currently the only independent composite index that covers closed and open models and refreshes every two weeks.
| # | Model | v5 | Percentile | GPQA | TB 4.0 | Input | Output | tok/s |
|---|---|---|---|---|---|---|---|---|
| 1 | Gemini 4 CarbonGoogle | 60.0 | 100% | — | 65.2 | $4.00 | $20.00 | 88 |
| 2 | Claude Opus 5.5Anthropic | 59.1 | 98% | 93.9 | 59.6 | $4.00 | $20.00 | 94 |
| 3 | GPT-6 AstraOpenAI | 56.9 | 95% | 95.8 | 59.1 | $10.00 | $50.00 | 81 |
| 4 | GPT-6.1 SolOpenAI | 54.6 | 91% | 95.4 | 56.1 | $2.00 | $10.00 | 77 |
| 5 | Claude Sonnet 5.5Anthropic | 54.5 | 91% | 95.6 | 63.6 | $2.00 | $10.00 | 135 |
| 6 | GPT-6.1 TerraOpenAI | 51.3 | 86% | 94.3 | 55.6 | $1.00 | $5.00 | 94 |
| 7 | Gemini 4 ArgonGoogle | 50.4 | 84% | 96.4 | 57.1 | $2.00 | $10.00 | 80 |
| 8 | Muse Wonder FullMeta | 50.0 | 83% | — | — | — | — | — |
| 9 | Claude Fable 5.1Anthropic | 49.0 | 82% | — | 52.0 | $10.00 | $50.00 | 57 |
| 10 | Kimi K3.1 Pro Open weightsMoonshot AI | 46.0 | 77% | — | 67.2 | $2.00 | $10.00 | 28 |
| 11 | Muse WonderMeta | 44.0 | 73% | — | — | $4.00 | $20.00 | 143 |
| 12 | Claude Haiku 5.5Anthropic | 36.0 | 60% | — | — | $0.10 | $0.50 | 235 |
Price × intelligence: who is overrated
Price analysis →The x-axis is blended price (3:1 input/output weighting, log scale); the y-axis is the intelligence index. Top-left is cheap and smart, bottom-right is expensive and not that smart.
Rankings by dimension
Methodology →A composite score hides real differences. The same model can match the flagship on document generation and score 1% on terminal agents — pick on the dimension you care about.
Reasoning
GPQA Diamond: graduate-level science questions that resist contamination
Code & terminal
Terminal-Bench 4.0: software-engineering tasks in real containers
Value
Intelligence points per dollar, at a 3:1 blended price
Output speed
tok/s throughput — how long long answers make you wait
Editor's notes
The five things in the last month of data that actually matter.
The intelligence index changed its ruler
The AA Intelligence Index moved to v5 on 27 October and every model's score shifted down. DeepSeek V4.1 Flash went from 45.1 on v4.3.2 to 33.0 on v5 — not because it got worse, but because the questions got harder. Comparing across versions is the most common misreading of rankings like this.
The middle of the market collapsed to $2/$10
GPT-6.1 Sol, Claude Sonnet 5.5, Kimi K3.1 Pro and Gemini 4 Argon now cost the same. Two years ago this tier was $15/$75. Price no longer signals which tier a model is in — only score and speed do.
Open weights take the agent crown
The day Kimi K3.1 Pro's weights opened (11-04), the top of Terminal-Bench 4.0 (67.2%) became an open-weights model; the best closed one, Gemini 4 Carbon, sits at 65.2%. Open models still trail by 3 points on knowledge — on getting work done they have already overtaken.
A 4.8× speed gap at the same price
Claude Sonnet 5.5 and Kimi K3.1 Pro both cost $2/$10, but output at 135 vs 28 tok/s. For a 4,000-character answer that is 30 seconds versus two and a half minutes.
Promo pricing has an expiry date
Gemini 4 Carbon's $2/$10 launch promo has ended and it now lists at $4/$20; Gemini 4 Argon's promo price doubles on 2026-12-01; Gemini 3.8 Flash's $0.75/$3.75 doubles on 2027-01-01. Budgeting at today's prices will mislead you.
You can't compute value per dollar for a subscription model
The newly catalogued Muse Wonder Full scores 50.0 on v5 — six points above the standard Wonder at 44.0, and six below its own v4.3.2 score of 56.0. But it is sold only inside the Muse Code subscription at $50 / $200 a month, with no token-priced API, so it enters the leaderboard and stays out of the price table. Comparing a subscription model with API models on points per dollar produces a number that means nothing.
Token usage
K-lines & full ranking →9,945 billion tokens network-wide on 2026-11-04, versus the previous session -0.17%. Read daily call volume like a stock: rising volume on the way up means share is being taken, falling volume on the way down means a model is being replaced.
AI directory
All 103 →The rankings answer "which model is best"; the directory answers "which one can I use right now". 103 sites in 12 categories, each flagged for free tier and overseas access.
C
C
G
D
D
P
M
CPick a model by use case
Tools →You need the strongest reasoning
Among closed models look at Claude Opus 5.5 (v5 59.1) and Gemini 4 Argon (GPQA 96.4). On the open side only Kimi K3 comes close at GPQA 93.1.
Reasoning ranking →You need it to change code
Terminal-Bench 4.0 is the only honest metric here. Any model below 20% should not go near a CI pipeline.
Coding ranking →You need the lowest possible cost
DeepSeek V4.1 Flash buys 62 intelligence points per dollar — 8.5× Opus 5.5 — under an MIT licence you can self-host.
Cost-efficiency ranking →