Degradation monitor
Degradation monitor v0.1
Raw measurements for one model across relays: in the same time slot every relay answers the same set of identity probes (knowledge cutoff, "the newest version you know of" and similar), and we record agreement with that model's reference answers, input tokens, reasoning tokens and the model fingerprint. The numbers are published as they are, with no verdict.
14 models, 101 routes
ModelRelaySort
| 100% (90 answers) | 109–24,260 | 1,014 | DeepSeek V4.1 Flash | — | 98.2% | 1.31 s | ¥2.59 | ||
| 100% (90 answers) | 109–24,260 | 1,014 | DeepSeek V4 Pro | — | 99.0% | 1.35 s | ¥11.4 | ||
| 100% (90 answers) | 109–24,260 | 1,014 | GLM 5.3 | — | 98.5% | 0.96 s | ¥8.56 | ||
| 100% (90 answers) | 109–24,260 | 1,014 | Qwen 3.8 Max | — | 98.0% | 1.06 s | ¥20.4 | ||
| 98% (90 answers) | 109–24,260 | 0 | Claude Haiku 5.5 | — | — | — | — | ||
| 98% (90 answers) | 100–161,349 | 38 | GPT 6.1 Sol | — | 98.7% | 0.87 s | ¥0.44 | ||
| 98% (90 answers) | — | 0 | Grok 4.7 | — | 98.3% | 5.11 s | ¥3.10 | ||
| 97% (90 answers) | 52–24,260 | 996 | Claude Fable 5.1 | — | 98.2% | 1.34 s | ¥4.38 | ||
| 97% (90 answers) | 52–24,260 | 996 | Claude Fable 5.1 | — | — | — | — | ||
| 97% (90 answers) | 52–24,260 | 996 | Claude Fable 5.1 | — | — | — | — | ||
| 97% (90 answers) | 52–24,260 | 996 | Claude Fable 5.1 | — | — | — | — | ||
| 97% (90 answers) | 52–24,260 | 559 | Claude Opus 5.5 | — | 97.3% | 1.57 s | ¥1.75 | ||
| 97% (90 answers) | 52–24,260 | 559 | Claude Opus 5.5 | — | — | — | — | ||
| 97% (90 answers) | 52–24,260 | 559 | Claude Opus 5.5 | — | — | — | — | ||
| 97% (90 answers) | 109–45,108 | 0 | Claude Sonnet 5 | — | 97.7% | 1.77 s | ¥7.30 | ||
| 97% (90 answers) | 109–45,108 | 0 | Claude Sonnet 5 | — | 97.7% | 1.58 s | ¥9.20 | ||
| 97% (90 answers) | 109–45,108 | 0 | Claude Sonnet 5 | — | 97.2% | 1.29 s | ¥0.88 | ||
| 97% (90 answers) | 109–45,108 | 0 | Claude Sonnet 5 | — | — | — | — | ||
| 97% (90 answers) | 109–24,260 | 0 | Claude Haiku 5.5 | — | 98.0% | 0.94 s | — | ||
| 97% (90 answers) | 109–24,260 | 0 | Claude Haiku 5.5 | — | — | — | — | ||
| 97% (90 answers) | 100–151,769 | 39 | GPT 6 Astra | — | 98.5% | 1.33 s | ¥2.19 | ||
| 97% (90 answers) | 100–151,769 | 39 | GPT 6 Astra | — | — | — | — | ||
| 97% (90 answers) | 100–151,769 | 39 | GPT 6 Astra | — | — | — | — | ||
| 97% (90 answers) | 100–151,769 | 39 | GPT 6 Astra | — | — | — | — | ||
| 97% (90 answers) | 100–151,769 | 39 | GPT 6 Astra | — | — | — | — | ||
| 97% (90 answers) | 100–161,349 | 38 | GPT 6.1 Sol | — | — | — | — | ||
| 97% (90 answers) | 100–161,349 | 38 | GPT 6.1 Sol | — | — | — | — | ||
| 97% (90 answers) | 100–161,349 | 38 | GPT 6.1 Sol | — | — | — | — | ||
| 97% (90 answers) | — | 0 | Grok 4.7 | — | — | — | — | ||
| 97% (90 answers) | — | 0 | Kimi K3 | — | 98.0% | 1.16 s | ¥28.6 | ||
| 85% (20 answers) | 101–22,953 | 23 | GPT 6 Astra | — | 100.0% | 9.76 s | — | ||
| 61% (18 answers) | 116–402 | 217 | Claude Opus 5.5 | — | 93.3% | 9.05 s | — | ||
| 44% (18 answers) | 116–139 | 243 | Claude Haiku 5.5 | — | 96.7% | 7.00 s | — | ||
| 39% (18 answers) | 114–411 | 0 | Claude Haiku 5.5 | — | 100.0% | 23.19 s | — | ||
| 69% (13 answers) | 3,685–21,590 | 0 | Gemini 3.8 Flash | — | — | — | — | ||
| 54% (28 answers) | 121–140 | 0 | Grok 4.7 | — | — | — | — | ||
| 28% (18 answers) | 239–748 | 0 | Claude Haiku 5.5 | — | 77.4% | 19.67 s | — | ||
| 100% (54 answers) | 456–694 | 69 | Claude Opus 5.5 | — | — | — | — | ||
| 100% (54 answers) | 472–730 | 178 | Claude Haiku 5.5 | — | — | — | — | ||
| 100% (8 answers) | 22,945–22,956 | 38 | GPT 6 Astra | — | 100.0% | 4.04 s | ¥3.30 | ||
| 96% (28 answers) | 5,264–5,279 | 29 | GPT 6 Astra | — | — | — | — | ||
| 89% (28 answers) | 4,203–4,218 | 27 | GPT 6 Astra | — | 100.0% | 8.02 s | — | ||
| 89% (28 answers) | 486–4,212 | 35 | GPT 6 Astra | — | — | — | — | ||
| 89% (18 answers) | 129–155 | 0 | Claude Fable 5.1 | — | — | — | — | ||
| 88% (8 answers) | 22,945–22,956 | 42 | GPT 6.1 Sol | — | — | — | — | ||
| 86% (28 answers) | 93–4,218 | 18 | GPT 6 Astra | — | — | — | — | ||
| 83% (18 answers) | 127–153 | 0 | Claude Sonnet 5 | — | 96.7% | 6.55 s | — | ||
| 83% (18 answers) | 116–139 | 0 | Claude Haiku 5.5 | — | — | — | — | ||
| 82% (28 answers) | 93–108 | 36 | GPT 6 Astra | — | 100.0% | 2.88 s | — | ||
| 82% (28 answers) | 812–828 | 0 | GPT 6.1 Sol | — | 100.0% | 3.49 s | ¥15.4 | ||
| 82% (28 answers) | 100–116 | 38 | GPT 6.1 Sol | — | — | — | — | ||
| 79% (28 answers) | 93–108 | 33 | GPT 6 Astra | — | 100.0% | 9.99 s | ¥2.04 | ||
| 79% (28 answers) | 93–108 | 34 | GPT 6.1 Sol | — | — | — | — | ||
| 72% (18 answers) | 129–155 | 537 | Claude Haiku 5.5 | — | 77.8% | 14.16 s | — | ||
| 68% (28 answers) | 4,207–4,480 | 16 | GPT 6.1 Sol | — | 96.7% | 5.20 s | — | ||
| 68% (28 answers) | 101–4,484 | 0 | GPT 6 Astra | — | 93.3% | 8.16 s | ¥11.9 | ||
| 61% (28 answers) | 4,211–18,563 | 0 | GPT 5.6 Terra | — | — | — | — | ||
| 56% (18 answers) | 116–139 | 0 | Claude Opus 5.5 | — | 83.3% | 12.04 s | ¥8.80 | ||
| 56% (18 answers) | 114–137 | 0 | Claude Haiku 5.5 | — | 77.8% | 9.30 s | ¥2.04 | ||
| 50% (18 answers) | 2,025–10,740 | 0 | Claude Haiku 5.5 | — | 79.3% | 21.85 s | — | ||
| 50% (18 answers) | 127–153 | 458 | Claude Sonnet 5 | — | 61.1% | 12.08 s | — | ||
| 44% (18 answers) | 116–139 | 170 | Claude Haiku 5.5 | — | 100.0% | 2.17 s | — | ||
| 44% (18 answers) | 258–2,860 | 0 | Claude Haiku 5.5 | — | 88.9% | 24.58 s | ¥2.04 | ||
| 44% (18 answers) | 153–168 | 0 | Claude Sonnet 5 | — | 100.0% | 3.97 s | ¥4.40 | ||
| 41% (17 answers) | 116–139 | 0 | Claude Haiku 5.5 | — | — | — | — | ||
| 40% (5 answers) | 137–144 | 0 | Claude Opus 5.5 | — | 60.0% | 2.38 s | ¥4.77 | ||
| 39% (18 answers) | 116–139 | 208 | Claude Haiku 5.5 | — | 100.0% | 2.03 s | — | ||
| 33% (18 answers) | 121–239 | 0 | Claude Haiku 5.5 | — | 100.0% | 2.88 s | ¥4.40 | ||
| 28% (18 answers) | 148–463 | 0 | Claude Haiku 5.5 | — | 88.9% | 15.94 s | ¥2.04 | ||
| Not measured yet | — | — | — | — | 100.0% | 23.19 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 23.19 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 8.85 s | — | ||
| Not measured yet | — | — | — | — | 98.6% | 1.40 s | ¥3.29 | ||
| Not measured yet | — | — | — | — | 98.2% | 0.91 s | ¥4.46 | ||
| Not measured yet | — | — | — | — | 73.3% | 1.83 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 2.27 s | — | ||
| Not measured yet | — | — | — | — | 77.4% | 3.02 s | ¥4.40 | ||
| Not measured yet | — | — | — | — | 100.0% | 3.52 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 6.04 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 2.45 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 4.62 s | ¥2.38 | ||
| Not measured yet | — | — | — | — | 100.0% | 7.18 s | ¥0.66 | ||
| Not measured yet | — | — | — | — | 90.0% | 6.70 s | — | ||
| Not measured yet | — | — | — | — | 86.7% | 4.67 s | — | ||
| Not measured yet | — | — | — | — | 0.0% | — | — | ||
| Not measured yet | — | — | — | — | 100.0% | 2.51 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 3.63 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 3.95 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 4.17 s | ¥1.30 | ||
| Not measured yet | — | — | — | — | 94.4% | 6.48 s | — | ||
| Not measured yet | — | — | — | — | 89.7% | 6.03 s | — | ||
| Not measured yet | — | — | — | — | 96.7% | 9.57 s | — | ||
| Not measured yet | — | — | — | — | 88.9% | 6.48 s | ¥2.04 | ||
| Not measured yet | — | — | — | — | 100.0% | 3.31 s | — | ||
| Not measured yet | — | — | — | — | 100.0% | 4.88 s | ¥1.27 | ||
| Not measured yet | — | — | — | — | 100.0% | 12.91 s | — | ||
| Not measured yet | — | — | — | — | 96.7% | 10.99 s | — | ||
| Not measured yet | — | — | — | — | 0.0% | — | — | ||
| Not measured yet | — | — | — | — | 96.7% | 4.34 s | — | ||
| Not measured yet | — | — | — | — | 88.9% | 5.53 s | ¥2.04 | ||
| Not measured yet | — | — | — | — | 90.0% | 7.27 s | — |
ProbesIn each slot every relay of a model gets the same set of identity probes, reworded and tagged with a random reference each time to avoid caches and targeted gaming
ComparisonEach answer is reduced to a short category and compared with the model's reference answers from a trusted route
MeasurementsProbe agreement, input tokens, median reasoning tokens, model fingerprint
PublicationRaw numbers are published directly, with every relay for the same model side by side
FAQ
How should these numbers be read?
Probes ask things like "when is your knowledge cutoff" or "which is the newest iPhone you know of": different models answer differently and one model answers consistently. Agreement well below other relays, input tokens that do not match the others, or a fingerprint unlike the model are worth a closer look. Single numbers move around; more answers make them more reliable.
Why look at input token counts?
The same input sent to the same model should always produce exactly the same token count. A count that differs from other relays usually means a different model, or extra content added to the request.
Why is there no data yet?
The degradation monitor is still being connected. The first measurements will appear here. Availability, time to first token and session cost in the table are live measurements.
How is this different from the main rankings?
The main rankings cover availability, speed and price. The degradation monitor shows whether each relay serves the real model. It does not affect rankings for now.