🌏 繁體中文版
Shift Summary
The Arena Agent leaderboard published a new batch on 2026-10-13, and third place changed again: Claude Fable 5.1 (Max) climbed back from sixth to third, while GPT-6 Sol (Max), which had only just entered the top three in the 10-12 batch, fell to seventh. First place, Claude Opus 5.5 (Max), held, and GPT-6.1 Sol (Max) kept second. Across the 10-08, 10-12 and 10-13 releases, third place went Fable 5.1, GPT-6 Sol, Fable 5.1 — and that back-and-forth is itself the signal worth recording today.
Ranking Changes
Arena Agent Leaderboard — 2026-10-13 (official release)
| Rank | Model | Score | 95% CI | 10-12 score (rank) | Change |
|---|---|---|---|---|---|
| 🥇 | Claude Opus 5.5 (Max) | 17.01% | [13.50%, 20.53%] | 17.89% (1) | — |
| 🥈 | GPT-6.1 Sol (Max) | 15.05% | [12.95%, 17.16%] | 13.81% (2) | — |
| 🥉 | Claude Fable 5.1 (Max) | 14.32% | [12.28%, 16.37%] | 11.74% (6) | ↑3 |
| 4 | GPT-6 Astra (Max) | 13.69% | [11.35%, 16.03%] | 12.34% (5) | ↑1 |
| 5 | Gemini 4 Argon (High) | 13.59% | [12.01%, 15.17%] | 12.37% (4) | ↓1 |
| 6 | Claude Sonnet 5.5 (Max) | 13.00% | [9.83%, 16.16%] | 10.75% (7) | ↑1 |
| 7 | GPT-6 Sol (Max) | 10.51% | [8.21%, 12.82%] | 12.60% (3) | ↓4 |
Source: Arena Leaderboard Dataset (agent subset, latest split) · official release date: 2026-10-13 · fetched: 2026-10-15 · 10-12 figures from the previous Benchmark Shift post
SWE-bench Verified's top five scores and ranks are unchanged from last time (79.2/79.2/78.8/77.4/76.8).
Analysis: What This Reshuffle Means
Technical
The gaps are tighter than in the last batch: second through sixth all sit between 13.00% and 15.05%, a spread of just 2.05 points. First-place Opus 5.5 (Max) again has among the fewest samples, 6,025 sessions, while fifth-place Gemini 4 Argon (High) has 26,905 — more than four times as many. Low-sample models naturally swing between batches, and Fable 5.1 (16,212 sessions) and GPT-6 Sol (6,664 sessions) each moved more than 2 points in two days.
Methodology
⚠️ Arena scores are relative win-rate scores, not Elo; across batches, only order and magnitude are meaningful.
The confidence intervals matter more: for second through sixth, the 95% CI lower bounds run from 9.83% to 12.95% and the upper bounds from 15.17% to 17.16%, overlapping heavily; even first place's lower bound of 13.50% falls inside every interval from second to sixth. Statistically, the order of these six models can't be distinguished. Third place changing three times in a week is exactly what random movement within overlapping intervals looks like, not models getting stronger or weaker.
Industry
For teams choosing a model, this table tells you which models sit in the same tier, not who's number one. Prices within that tier vary a lot: Claude Opus 5.5 is $4/$20 per million tokens, GPT-6.1 Sol is $2/$10, and Gemini 4 Argon is $2/$10 through November 30 and $4/$20 from December 1. When scores can't be separated statistically, testing on your own tasks and comparing cost per task is more reliable than chasing leaderboard rank.
Today's Takeaway
In the last post I read GPT-6 Sol's jump into the top three as "this batch happened to suit its strengths" — and a day later it fell back to seventh. That corrected how I read this board: on a leaderboard with overlapping confidence intervals, a single batch's rank jump doesn't need an explanation; a model holding the same position over several batches is what's worth explaining.