Rank01Proprietary
Claude Fable 5
Anthropic
- Reasoning index #3
- Text preference #1
- Agent tasks #2
96
100%
3 benchmark signals
LLM Leaderboard
A traceable LLM ranking snapshot that brings together multiple public benchmark leaderboards into one decision table, refreshed every day.
Models
20
Signals
48
Updated
2026-10-23
Refreshed daily at 00:00 HKT
Consensus snapshot
Major benchmark signals, normalized into one decision table.
#1Claude Fable 5Anthropic · Reasoning index #3 · Text preference #1 · Agent tasks #296
#2Claude Fable 5.1Anthropic · Reasoning index #194
#3Claude Opus 5Anthropic · Reasoning index #2 · Text preference #6 · Agent tasks #192
#4Claude Opus 4.7Anthropic · Text preference #3 · Agent tasks #879
Consensus Ranking
The consensus score is not a raw benchmark score. It normalizes cross-benchmark positions and groups effort levels and variants of one product line under its best result.
Some benchmark sources could not be refreshed; those signals reuse the last successful snapshot.
Weight mixReasoning index 40%Text preference 30%Agent tasks 30%
RankModelScoreCoverage
Rank02Proprietary
Claude Fable 5.1
Anthropic
- Reasoning index #1
94
40%
1 benchmark signals
Rank03Proprietary
Claude Opus 5
Anthropic
- Reasoning index #2
- Text preference #6
- Agent tasks #1
92
100%
3 benchmark signals
Rank04Proprietary
Claude Opus 4.7
Anthropic
- Text preference #3
- Agent tasks #8
79
60%
2 benchmark signals
Rank05Proprietary
GPT-5.6 Sol
OpenAI
- Reasoning index #4
- Text preference #13
- Agent tasks #3
78
100%
3 benchmark signals
Rank06Open weights
Kimi K3
Kimi
- Reasoning index #6
- Text preference #9
- Agent tasks #5
78
100%
3 benchmark signals
Rank07Proprietary
Gemini 3.8 Flash
- Reasoning index #8
- Text preference #5
- Agent tasks #11
72
100%
3 benchmark signals
Rank08Proprietary
Claude Opus 4.6
Anthropic
- Text preference #2
- Agent tasks #15
67
60%
2 benchmark signals
Rank09Proprietary
Claude Opus 4.8
Anthropic
- Text preference #16
- Agent tasks #4
61
60%
2 benchmark signals
Rank10Proprietary
GPT 5.5
OpenAI
- Text preference #14
- Agent tasks #6
61
60%
2 benchmark signals
Rank11Open weights
Qwen3.8 2.4T A95B
Alibaba
- Reasoning index #10
60
40%
1 benchmark signals
Rank12Proprietary
Muse Spark
Meta
- Text preference #10
60
30%
1 benchmark signals
Rank13Open weights
GLM-5.3
Z AI
- Reasoning index #7
- Text preference #15
- Agent tasks #17
54
100%
3 benchmark signals
Rank14Proprietary
Gemini 3 Pro
- Text preference #12
52
30%
1 benchmark signals
Rank15Proprietary
Qwen3.8 Max
Alibaba
- Reasoning index #9
- Text preference #18
- Agent tasks #14
51
100%
3 benchmark signals
Rank16Proprietary
Muse Spark 1.2
Meta
- Reasoning index #12
- Text preference #4
- Agent tasks #25
50
100%
3 benchmark signals
Rank17Proprietary
Grok 4.6
SpaceXAI
- Reasoning index #5
- Text preference #40
- Agent tasks #13
49
100%
3 benchmark signals
Rank18Proprietary
Gemini 3.7 Flash
- Reasoning index #14
- Text preference #8
- Agent tasks #27
41
100%
3 benchmark signals
Rank19Proprietary
Claude Sonnet 5
Anthropic
- Reasoning index #17
- Text preference #39
- Agent tasks #7
37
100%
3 benchmark signals
Rank20Open weights
GLM-5.3-Flash
Z AI
- Reasoning index #11
- Text preference #25
- Agent tasks #16
37
100%
3 benchmark signals
Methodology
Weighted consensus, not benchmark shopping.
- 01Three public benchmark leaderboards are fetched automatically every day at 00:00 HKT and cached for 24 hours.
- 02Each benchmark rank is converted into position points, then averaged with the declared weight mix.
- 03Effort levels and variants of one product line are grouped, keeping only its best result per benchmark.
- 04Models missing some benchmark coverage receive a small coverage discount, so a single leaderboard spike does not dominate the table.
- 05The ranking is a procurement and experimentation entry point, not an absolute answer. Cost, latency, context, data policy, and task-level testing still matter.
Turn rankings into a practical model selection process.
If you want to connect this leaderboard to model selection, prompt tests, or agent workflows, we can turn it into a repeatable evaluation lane.