Meta Muse Wonder (max) at 162 tokens per second
The Intelligence Index, full top fifteen
| # | Model configuration | Provider | Index | Median tok/s |
|---|
| 1 | Claude Opus 5.5 (max, fallback) | Anthropic | 58 | 96 |
| 2 | Claude Sonnet 5.5 (max, fallback) | Anthropic | 56 | 134 |
| 3 | Claude Opus 5.5 (xhigh, fallback) | Anthropic | 56 | 82 |
| 4 | Muse Wonder Full (max) | Meta | 56 | not stated |
| 5 | Kimi K3.1 Pro | Moonshot AI | 55 | 58 |
| 6 | Claude Opus 5.5 (high, fallback) | Anthropic | 54 | 80 |
| 7 | Claude Fable 5.1 (max, fallback) | Anthropic | 53 | 70 |
| 8 | Claude Fable 5.1 (xhigh, fallback) | Anthropic | 53 | 63 |
| 9 | GPT-6 Astra (max) | OpenAI | 53 | 66 |
| 10 | Gemini 4 Argon (high) | Google | 53 | 81 |
| 11 | Muse Wonder (max) | Meta | 52 | 162 |
| 12 | GPT-6 Astra (xhigh) | OpenAI | 52 | 62 |
| 13 | Claude Sonnet 5.5 (xhigh, fallback) | Anthropic | 52 | 108 |
| 14 | GPT-6.1 Sol (max) | OpenAI | 52 | 59 |
| 15 | Claude Opus 5.5 (medium, fallback) | Anthropic | 51 | 83 |
Source: the Artificial Analysis Intelligence Index leaderboard, read 26 October 2026. AIToolsRecap has not
benchmarked any of these models independently.
Read the configuration, not the model name
The single most useful thing on that table is the parenthesis. Claude Opus 5.5 appears four times - at max,
xhigh, high and medium reasoning effort - and spans index 58 down to 51. That is a seven-point range inside one
model. GPT-6 Astra appears twice, at 53 and 52. Someone telling you they are using Opus 5.5 has told you almost
nothing about what they are actually paying for or how well it performs.
The word fallback matters just as much. Most Anthropic entries carry it, meaning the
configuration is permitted to retry on a different setting. A score produced with fallback and a score produced
without are not directly comparable, and that distinction is invisible if you read the table as a simple list of
model names.
The new entrants
Meta Muse Wonder Full entering at fourth is the bigger surprise of the two, because Meta had no
presence in October's top ten at all. Its sibling, Muse Wonder (max), sits eleventh on 52 and is the fastest
configuration in the entire top fifteen at 162 tokens per second - roughly double the Claude Opus 5.5
configurations around it.
Kimi K3.1 Pro at fifth is the first non-US model to reach the top five on this board. It runs at
58 tokens per second, which is near the bottom of the range here, so it is buying its position with quality
rather than speed.
GPT-6.1 Sol: the one model with published prices
The Intelligence Index publishes cost per task rather than per-token pricing, which makes direct price comparison
awkward. GPT-6.1 Sol is the exception this month because OpenAI published its rates directly:
$2.00 per million input tokens, $10.00 per million output tokens, and $0.10 for cached input -
the cached rate being half what GPT-6 Sol charged.
OpenAI's own figures put it level with GPT-6 Astra on DeepSWE v1.1 at roughly one-fifth the cost, 6.4 percentage
points above GPT-6 Sol's best score on the same benchmark, and 2.2 points above Claude Opus 5.5 on
AutomationBench 1.0.6 at medium reasoning effort. Its factual error rate at low effort falls from GPT-6 Sol's
11.4% to 7.7%. These are vendor-published comparisons, which is worth holding in mind - it sits fourteenth on the
independent index at 52.
Reading any leaderboard without being misled
Six checks, in the order they matter:
1. Check the configuration in parentheses before comparing anything.
2. Check whether fallback is enabled - it changes what the score means.
3. Check the date the board was read. These move weekly.
4. Check whether a figure is measured or estimated by the board itself.
5. Separate vendor-published benchmarks from independent ones.
6. Check cost per task, not index position, if you are paying the bill.
Frequently asked questions
What is the best AI model in November 2026?
By the Artificial Analysis Intelligence Index, Claude Opus 5.5 at max reasoning effort with fallback, scoring 58. Meta Muse Wonder Full and Claude Sonnet 5.5 tie behind it on 56.
Did Claude lose the top spot?
No. Claude Opus 5.5 still holds first on 58. What changed is the top five - it was entirely Anthropic in October and now includes Meta at fourth and Kimi K3.1 Pro at fifth.
What is Muse Wonder Full?
Meta's model, which entered the Intelligence Index at fourth with a score of 56. Its sibling Muse Wonder (max) sits eleventh on 52 and is the fastest configuration in the top fifteen at 162 tokens per second.
How much does GPT-6.1 Sol cost?
OpenAI publishes $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.10 per million - half GPT-6 Sol's cached rate.
Why does the same model appear several times?
Because reasoning effort is scored separately. Claude Opus 5.5 appears four times, from 58 at max effort down to 51 at medium. The configuration matters more than the model name.
Is Kimi K3.1 Pro the first Chinese model in the top five?
It is the first non-US model to reach the top five on this board. It runs at 58 tokens per second, near the bottom of the range for the top fifteen.
Sources
Index scores and speeds from the
Artificial Analysis Intelligence Index,
read 26 October 2026. GPT-6.1 Sol pricing and benchmark comparisons from OpenAI's own announcement,
Introducing GPT-6.1 Sol.
October comparison figures from our own
October 2026 rankings.
Rankings as published by a third-party board on 26 October 2026.
AIToolsRecap has not independently benchmarked any model listed. Last updated 26 October 2026.