Leaderboard
Preference ratings from blind human votes, on an Elo-like scale — cost and speed were never visible to voters.
Kimi K3
1580
±144
9/27 picked · $0.424/gen
provisional
Claude Fable 5
1551
±146
8/28 picked · $1.04/gen
provisional
GPT 5.6 Sol
1501
±154
6/28 picked · $0.305/gen
provisional
| # | Model | Rating | W / L | Votes | Avg cost | Median time |
|---|---|---|---|---|---|---|
| 1▲2 | Kimi K3Moonshotprovisional | 1580±144 | 27 | $0.424 | ||
| 2▼1 | Claude Fable 5Anthropicprovisional | 1551±146 | 28 | $1.04 | ||
| 3▼1 | GPT 5.6 SolOpenAIprovisional | 1501±154 | 28 | $0.305 | ||
| 4 | Claude Opus 5Anthropic | 1500 | 0 | $1.26 | ||
| 5▼1 | GLM 5.2Z.aiprovisional | 1495±153 | 29 | $0.066 |
Ratings come from one preference model fitted to every blind vote at once (Plackett–Luce, on an Elo-like scale where +400 ≈ 10:1 odds). Because it’s fitted globally rather than accumulated vote by vote, a model added later is judged on its own record — not penalised for arriving late. ± is the 95% confidence interval; overlapping intervals mean the gap isn’t established yet. “Provisional” means too few votes to place confidently. W/L is times picked vs times shown. Cost and time are per full generation. ▲/▼ is the 7-day rank move.
Rating vs cost
Up and left is the value frontier — preferred by humans, cheap to run. The leader is highlighted in gold.
Bubble size reflects vote volume.