Biggest AI model ranking changes — October 26, 2026: a LiveBench refresh reshuffles the League.
The biggest AI model ranking changes in today's League table did not come from a launch — they came from a benchmark refresh. Yesterday's and today's snapshots of our own database, diffed model by model, show one model down twenty points, a new leader at the top, and several models that rose without their own scores improving at all.
By AI Pro League · Published

These AI model ranking changes are computed from two files this site publishes itself: the League standings snapshots of Nov 9, 2026 and Nov 10, 2026. Of the 45 scored models present in both, 20 moved by two points or more and 15 were effectively unchanged. No model was added or removed overnight — every move below is a score move, not a line-up change.
The biggest fallers: AI model ranking changes at the bottom
| Model | Maker | League Score | Change | Rank |
|---|---|---|---|---|
| GPT-6.1 Sol | OpenAI | 96.6 → 93.3 | -3.3 | 4 → 6 |
| Claude Sonnet 5.5 | Anthropic | 87.9 → 85.0 | -2.9 | 6 → 11 |
| Muse Wonder | Meta | 75.5 → 73.0 | -2.5 | 12 → 18 |
| GPT-6 Luna | OpenAI | 55.1 → 53.3 | -1.8 | 30 → 34 |
| Claude Opus 5.5 | Anthropic | 98.7 → 97.1 | -1.6 | 2 → 3 |
The headline fall is GPT-6.1 Sol: 96.6 → 93.3 (-3.3), and rank 4 → 6. It is not a mystery move — the model’s only recorded component is LiveBench, and its raw LiveBench score was re-measured on Nov 10, 2026 at 81.4 against 84.1 the day before. When a model carries a single component, a single re-measurement moves its whole League Score: the normalised component fell from 96.6 to 93.3, and the League Score followed point for point.
The biggest risers
| Model | Maker | League Score | Change | Rank |
|---|---|---|---|---|
| Muse Spark 1.3 | Meta | 51.7 → 68.6 | +16.9 | 32 → 25 |
| Inkling | Thinking Machines Lab | 24.6 → 36.1 | +11.5 | 39 → 41 |
| Claude Haiku 4.5 | Anthropic | 26.9 → 33.8 | +6.9 | 38 → 42 |
| Kimi K3.1 Pro | Moonshot AI | 69.4 → 75.4 | +6.0 | 18 → 14 |
| Nemotron 3 Ultra 550B-A55B | NVIDIA | 32.3 → 37.8 | +5.5 | 36 → 40 |
The most instructive rise is Claude Haiku 4.5 (+6.9, to 33.8). Its recorded LiveBench overall raw score did not change at all — 46.0 on both days. What changed was the field around it: League Score components are normalised across the models we track, so when other models’ fresh scores reshaped the range, this model’s normalised component rose from 15.5 to 25.9 without the model doing anything. A League Score is a position in a field, not a fixed grade — today proves it in both directions.
Why one benchmark refresh moves the whole table
LiveBench carries 40% of the League Score — the largest single weight in the League Score methodology — and it re-measures models on fresh, contamination-resistant questions (the project’s own site is LiveBench). A refresh therefore does two things at once: it changes the raw scores of the models it re-tested, and it changes the normalisation range every other model is scaled against. That is why 20 of 45 scored models could move by two points or more on a day with no launches at all.
How to read a one-day move
- Check the coverage first. A model scored on one component (like the biggest faller above) will always swing harder than a model averaged over four or five — its page shows exactly which components it carries.
- Separate raw from normalised. A raw change means the model was re-measured; a normalised-only change means the field moved. The two mean different things, and each model page shows both.
- One day is weather, not climate. A single refresh can reverse at the next one. The moves worth acting on are the ones that survive a second snapshot — which is why we publish the daily data instead of a headline.
See the table these moves produced on the current League standings, compare any two movers head-to-head on the model comparison hub, and read how the normalisation itself works in our GPQA Diamond explainer.