Change Significance Tracker
Not all rank changes are meaningful. Some are random noise. This page uses statistical analysis to tell you which model score movements are real trends vs. normal fluctuation, so you know which changes to pay attention to.
Models Analyzed
285
Significant Changes
19
Noise (Not Significant)
266
Both Timeframes
18
What This Means
- 19 of 285 models have score changes that are statistically significant - these are real performance shifts, not random noise.
- 266 models show score changes within their normal variation range - don't read too much into small rank shifts for these models.
- 18 models show significant changes on both daily and weekly timeframes - these are the strongest, most reliable signals of real performance change.
- How to use this: When a model's rank changes, check here first. If it's not flagged as significant, the change is likely temporary noise. If it is significant, the model is genuinely improving or declining.
Top Significant Changes (by |Z-Score|)
Real Performance Shifts
19 models whose recent scores deviate enough from their historical average to be considered a real change (not noise). Sorted by how extreme the change is. Z-Score measures how unusual the change is - values beyond ±1.96 mean there's a 95% chance the change is real.
| Model | Current | Z-Score | Direction |
|---|---|---|---|
| Claude Sonnet 5Anthropic | 85.2 | 4.24 | Improvement |
| Claude Opus 5Anthropic | 95.1 | 3.87 | Improvement |
| Grok 4.6xAI | 88.8 | 3.46 | Improvement |
| GLM 5.1Zhipu AI | 68.3 | -2.98 | Decline |
| Gemma 4 31B (free)Google | 75.9 | -2.96 | Decline |
| GPT-5 MiniOpenAI | 40.5 | -2.72 | Decline |
| GPT-4o (2024-11-20)OpenAI | 71.2 | 2.52 | Improvement |
| GPT-5 NanoOpenAI | 76.8 | 2.48 | Improvement |
| Grok 4.7xAI | 88.8 | 2.45 | Improvement |
| Claude Opus 5.5Anthropic | 95.1 | 2.45 | Improvement |
| GPT-4.1 MiniOpenAI | 44.2 | -2.36 | Decline |
| Claude Sonnet 5.5Anthropic | 85.2 | 2.24 | Improvement |
| GLM 5.2Zhipu AI | 68.9 | -2.20 | Decline |
| Hy3Tencent | 40.0 | -2.13 | Decline |
| Muse Spark 1.1meta | 40.0 | -2.08 | Decline |
| Inklingthinkingmachines | 40.0 | -2.08 | Decline |
| Kimi K2.6Moonshot AI | 47.5 | -2.01 | Decline |
| Claude Haiku 5.5Anthropic | 69.9 | 2.00 | Improvement |
| Gemini 3.5 Flash LiteGoogle | 79.0 | 1.98 | Improvement |
Short-Term vs. Sustained Changes
A model changing rank in 24 hours could be a blip. But if it's also moving over 7 days, that's a real trend. Models flagged on both timeframes are the most important to watch - they represent confirmed, sustained performance shifts.
Significant on Both Timeframes(strongest signals)
| Model | Score | 24h Change | 7d Change |
|---|---|---|---|
| Hy3Tencent | 40.0 | -6 | -16 |
| Muse Spark 1.1meta | 40.0 | -6 | -16 |
| Inklingthinkingmachines | 40.0 | -6 | -17 |
| Inkling Smallthinkingmachines | 40.0 | -5 | -18 |
| Gemini 3.6 FlashGoogle | 40.0 | -6 | -17 |
| Kimi K3Moonshot AI | 40.0 | -6 | -16 |
| Gemini 3.6 Flash (batch)Google | 40.0 | -6 | -17 |
| Muse Spark 1.2meta | 40.0 | -4 | -17 |
| DeepSeek V4 Flash Latest~deepseek | 40.0 | -4 | -17 |
| DeepSeek V4 Flash 0731DeepSeek | 40.0 | -4 | -17 |
| Qwen3.7 FlashAlibaba | 40.0 | -6 | -18 |
| Claude Opus 5 (batch)Anthropic | 40.0 | -6 | -17 |
| Ling 3.0 Flashinclusionai | 40.0 | -6 | -17 |
| LongCat 2.0Meituan | 40.0 | -6 | -17 |
| Kimi K3 (batch)Moonshot AI | 40.0 | -6 | -16 |
| Grok Latest~x-ai | 40.0 | -6 | -16 |
| Aion-3.0-Miniaion-labs | 40.0 | -6 | -16 |
| Aion-3.0aion-labs | 40.0 | -6 | -16 |
Daily Only(may be noise)
No models with daily-only significance.
Weekly Only(building trend)
Which Models Are Noisy vs. Consistent?
Some models have naturally stable scores - even a small rank change for these models is meaningful. Others have volatile scores that bounce around - they need a bigger shift before you should care. CV% (coefficient of variation) tells you how volatile each model is. Higher = noisier.
Noisiest Models(highest CV% - widest significance thresholds)
| Model | Score | CV% |
|---|---|---|
| Claude Opus 5.5Anthropic | 95.1 | 40.3% |
| Gemma 2 27BGoogle | 77.4 | 37.5% |
| Claude Sonnet 5.5Anthropic | 85.2 | 35.4% |
| Phi 4Microsoft | 61.0 | 34.2% |
| Gemini 3.6 FlashGoogle | 40.0 | 33.3% |
| MiMo-V2.6-Pro-UltraSpeedXiaomi | 40.0 | 32.5% |
| MiMo-V2.6-ProXiaomi | 40.0 | 32.5% |
| Step 5 PreviewStepFun | 40.0 | 30.8% |
| Claude Opus 5Anthropic | 95.1 | 30.7% |
| MiMo-V2.6-FlashXiaomi | 40.0 | 30.6% |
| Kimi K3Moonshot AI | 40.0 | 28.8% |
| GPT-4OpenAI | 60.0 | 28.7% |
| Mixtral 8x22B InstructMistral AI | 63.4 | 28.3% |
| Llama 3.3 70B InstructMeta | 66.8 | 28.1% |
| Mistral LargeMistral AI | 65.9 | 28.0% |
| Llama 3.1 70B InstructMeta | 63.6 | 27.9% |
| DeepSeek V3DeepSeek | 69.5 | 26.9% |
| Command A+Cohere | 53.2 | 26.6% |
| Command ACohere | 51.6 | 26.6% |
| Mistral Large 2407Mistral AI | 65.9 | 26.2% |
Most Consistent Models(lowest CV% - tightest significance thresholds)
| Model | Score | CV% |
|---|---|---|
| GPT-5.2 Pro (batch)OpenAI | 90.5 | 0.0% |
| GPT-5.6 Luna Pro (batch)OpenAI | 89.0 | 0.0% |
| GPT-5.6 Luna (batch)OpenAI | 89.0 | 0.0% |
| GPT-5.6 Terra Pro (batch)OpenAI | 89.0 | 0.0% |
| GPT-5.6 Terra (batch)OpenAI | 89.0 | 0.0% |
| GPT-5.6 Sol Pro (batch)OpenAI | 89.0 | 0.0% |
| GPT-5.6 Sol (batch)OpenAI | 89.0 | 0.0% |
| Gemini 3 Flash Preview (batch)Google | 88.4 | 0.0% |
| Grok 4.20 Multi-AgentxAI | 87.9 | 0.0% |
| Gemini 3.5 Flash (batch)Google | 79.0 | 0.0% |
| GPT-4o (batch)OpenAI | 72.0 | 0.0% |
| GPT-4o-mini (batch)OpenAI | 62.5 | 0.0% |
| Claude Haiku 5.5 (batch)Anthropic | 58.0 | 0.0% |
| Kimi K3.1 ProMoonshot AI | 40.0 | 0.0% |
| Flint (free)recursive | 40.0 | 0.0% |
| MiniMax M3.1 FlashMiniMax | 40.0 | 0.0% |
| DeepSeek V4.1 Flash 1017DeepSeek | 40.0 | 0.0% |
| Glimpseelorian | 40.0 | 0.0% |
| Xing4.0 29B A4B (free)xingchen-agi | 40.0 | 0.0% |
| Beam 1.0 501B A23Breflection | 40.0 | 0.0% |
How Significance Is Calculated
Understanding the statistical methodology behind our significance analysis helps you distinguish real performance shifts from random fluctuations.
Statistical Significance
We use z-scores with a 95% confidence threshold (|z| > 1.96). A z-score measures how many standard deviations a model's current score is from its historical baseline. Only changes exceeding 1.96 standard deviations are flagged as statistically significant.
Baseline Score
The baseline is computed as the arithmetic mean of each model's 14-day sparkline data. This rolling average smooths out daily fluctuations and provides a stable reference point for detecting meaningful deviations.
Confidence Intervals
Each model's 95% confidence interval is calculated as baseline ± 1.96 × standard deviation. Scores falling outside this range indicate a statistically meaningful change. The "Confidence" column shows the ± threshold value.
Multi-Timeframe Analysis
Daily (24h) and weekly (7d) rank changes are analyzed separately. Daily significance requires a rank shift of more than 3 positions; weekly requires more than 5. Models significant on both timeframes represent the strongest, most reliable signals.
Noise vs. Signal
The coefficient of variation (CV%) measures relative volatility. High-CV models have naturally noisy scores and require larger absolute changes to be significant. Low-CV models are more predictable, so even small deviations may represent real shifts.
Related
Statistical significance indicates whether a model's rank change represents a real performance shift or is just random noise. We use z-scores with a 95% confidence threshold (|z| > 1.96), meaning a change is only flagged as significant if there is less than a 5% chance it occurred by random variation.
A z-score measures how many standard deviations a model's current score deviates from its historical baseline. It is calculated as (current score - baseline mean) / standard deviation. Values above +1.96 indicate significant improvement, while values below -1.96 indicate significant decline.
The CV% measures a model's relative score volatility. A high CV% means the model's performance fluctuates a lot, requiring larger changes to be statistically significant. A low CV% means the model is very consistent, so even small deviations may represent meaningful shifts. This helps distinguish inherently noisy models from truly changing ones.