Live model degradation tracker
AI degradation tracker Is any model getting worse right now?
Every model on the board is re-tested on the same tasks, coding every four hours, reasoning and tool use daily, with an hourly canary probe. Each suite has its own drift detector against the model’s own past, the canary watches for sharp drops, and when either raises an alarm the Smart Router routes around that model in the modes the alarm affects.
Three models free · email when a watched model's coding score falls five points in a week
1,000 routed requests a month free · your own provider keys · no card
As of 30 Oct 2026, 16:16 UTC · 25 models · 0 open alarms
- 0
- models have an open alarm, out of 25 on the tracker
- 66 of 75
- per-suite detectors armed with no detection
- 3
- canary alarms in the last 30 days, each one logged below
01
Is any AI model getting worse right now?
None of the 25 models on the tracker has an open alarm. 66 of 75 per-suite detectors are armed with no detection, and 3 models are still calibrating.
As of 30 Oct 2026, 16:16 UTC · per-suite Page-Hinkley detectors, the 24-hour boards and the hourly canary · the page refreshes every 15 minutes
02
Every model, suite by suite
Each pill is the detector’s verdict on one suite. The bar under it shows how close the detector’s statistic is to the alarm threshold, and the score beside it comes from the 24-hour board.
GPT degradation tracker
None of the 9 GPT models on the tracker has an open alarm.
| Model | Coding | Reasoning | Tool use | Canary, hourly |
|---|---|---|---|---|
| GPT-6 Luna | No declinedetector at 0.00 of 0.3024-hour score 94 | No declinedetector at 0.00 of 0.3024-hour score 98 | No declinedetector at 0.00 of 0.5024-hour score 88 | No alarmnone in 30 days |
| GPT-5.6 Luna | No declinedetector at 0.01 of 0.3024-hour score 93 | No declinedetector at 0.00 of 0.3024-hour score 94 | No declinedetector at 0.01 of 0.5024-hour score 86 | No alarmnone in 30 days |
| GPT-6.1 Sol | No declinedetector at 0.00 of 0.3024-hour score 93 | No declinedetector at 0.00 of 0.3024-hour score 95 | No declinedetector at 0.00 of 0.5024-hour score 85 | No alarmnone since tracking began |
| GPT-5.6 Terra | No declinedetector at 0.00 of 0.3024-hour score 94 | No declinedetector at 0.00 of 0.3024-hour score 97 | No declinedetector at 0.00 of 0.5024-hour score 80 | No alarmnone in 30 days |
| GPT-5.6 Sol | No declinedetector at 0.05 of 0.3024-hour score 91 | No declinedetector at 0.00 of 0.3024-hour score 99 | No declinedetector at 0.00 of 0.5024-hour score 83 | No alarmnone in 30 days |
| GPT-6 Sol | No declinedetector at 0.02 of 0.3024-hour score 91 | No declinedetector at 0.00 of 0.3024-hour score 96 | No declinedetector at 0.00 of 0.5024-hour score 84 | No alarmnone in 30 days |
| GPT-6 Astra | No declinedetector at 0.02 of 0.3024-hour score 91 | No declinedetector at 0.00 of 0.3024-hour score 93 | No declinedetector at 0.00 of 0.5024-hour score 86 | No alarmnone in 30 days |
| GPT-5.5 | No declinedetector at 0.01 of 0.3024-hour score 93 | No declinedetector at 0.00 of 0.30 · community-funded runs24-hour score 58 | No declinedetector at 0.01 of 0.50 · community-funded runs24-hour score 76 | No alarmnone in 30 days |
| GPT-6.1 TerraCalibrating | Calibratingday 5 of 1024-hour score 91 | Calibratingday 5 of 1024-hour score 93 | Calibratingday 5 of 1024-hour score 79 | No alarmnone since tracking began |
Claude degradation tracker
None of the 9 Claude models on the tracker has an open alarm.
| Model | Coding | Reasoning | Tool use | Canary, hourly |
|---|---|---|---|---|
| Claude Sonnet 5.5 | No declinedetector at 0.00 of 0.3024-hour score 95 | No declinedetector at 0.00 of 0.3024-hour score 100 | No declinedetector at 0.00 of 0.5024-hour score 90 | No alarmnone since tracking began |
| Claude Opus 5.5 | No declinedetector at 0.00 of 0.3024-hour score 97 | No declinedetector at 0.00 of 0.3024-hour score 98 | No declinedetector at 0.00 of 0.5024-hour score 85 | No alarmnone in 30 days |
| Claude Fable 5.1 | No declinedetector at 0.25 of 0.3024-hour score 89 | No declinedetector at 0.00 of 0.3024-hour score 100 | No declinedetector at 0.03 of 0.5024-hour score 88 | No alarmnone in 30 days |
| Claude Opus 5 | No declinedetector at 0.00 of 0.3024-hour score 88 | No declinedetector at 0.00 of 0.3024-hour score 97 | No declinedetector at 0.00 of 0.5024-hour score 90 | No alarmnone in 30 days |
| Claude Sonnet 5 | No declinedetector at 0.00 of 0.3024-hour score 90 | No declinedetector at 0.00 of 0.3024-hour score 91 | No declinedetector at 0.04 of 0.5024-hour score 85 | No alarmlast on 24 Oct 2026, closed after 3 h |
| Claude Opus 4.7 | No declinedetector at 0.00 of 0.3024-hour score 94 | No declinedetector at 0.00 of 0.30 · community-funded runs24-hour score 51 | No declinedetector at 0.08 of 0.50 · community-funded runs24-hour score 70 | No alarmnone in 30 days |
| Claude Opus 4.8 | No declinedetector at 0.00 of 0.3024-hour score 85 | No declinedetector at 0.00 of 0.30 · community-funded runs24-hour score 55 | No declinedetector at 0.14 of 0.50 · community-funded runs24-hour score 77 | No alarmnone in 30 days |
| Claude Opus 4.6 | No declinedetector at 0.00 of 0.3024-hour score 91 | No declinedetector at 0.00 of 0.30 · community-funded runs24-hour score 49 | No declinedetector at 0.10 of 0.50 · community-funded runs24-hour score 67 | No alarmnone in 30 days |
| Claude Sonnet 4.6 | No declinedetector at 0.00 of 0.3024-hour score 90 | No declinedetector at 0.00 of 0.30 · community-funded runs24-hour score 46 | No declinedetector at 0.29 of 0.50 · community-funded runs24-hour score 67 | No alarmnone in 30 days |
Is Claude getting worse?Every Claude modelAnthropic API status
Gemini degradation tracker
None of the 3 Gemini models on the tracker has an open alarm.
| Model | Coding | Reasoning | Tool use | Canary, hourly |
|---|---|---|---|---|
| Gemini 3.8 Flash | No declinedetector at 0.00 of 0.3024-hour score 92 | No declinedetector at 0.00 of 0.3024-hour score 100 | No declinedetector at 0.00 of 0.5024-hour score 78 | No alarmnone in 30 days |
| Gemini 3.5 Flash Lite | No declinedetector at 0.00 of 0.3024-hour score 87 | No declinedetector at 0.00 of 0.3024-hour score 100 | No declinedetector at 0.00 of 0.5024-hour score 79 | No alarmlast on 25 Oct 2026, closed after 2 h |
| Gemini 4 ArgonCalibrating | Calibratingday 5 of 1024-hour score 91 | Calibratingday 5 of 1024-hour score 97 | Calibratingday 5 of 1024-hour score 87 | No alarmlast on 28 Oct 2026, closed after 1 h |
DeepSeek degradation tracker
None of the 2 DeepSeek models on the tracker has an open alarm.
| Model | Coding | Reasoning | Tool use | Canary, hourly |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | No declinedetector at 0.02 of 0.3024-hour score 94 | No declinedetector at 0.00 of 0.3024-hour score 97 | No declinedetector at 0.00 of 0.5024-hour score 88 | No alarmnone in 30 days |
| DeepSeek V4 Pro | No declinedetector at 0.00 of 0.3024-hour score 86 | No declinedetector at 0.00 of 0.3024-hour score 95 | No declinedetector at 0.00 of 0.5024-hour score 92 | No alarmnone in 30 days |
Is DeepSeek getting worse?Every DeepSeek modelDeepSeek API status
Kimi degradation tracker
The Kimi model on the tracker has no open alarm.
| Model | Coding | Reasoning | Tool use | Canary, hourly |
|---|---|---|---|---|
| Kimi K3.1 ProCalibrating | Calibratingday 5 of 1024-hour score 87 | Calibratingday 5 of 1024-hour score 94 | Calibratingday 5 of 1024-hour score 86 | No alarmnone since tracking began |
GLM degradation tracker
The GLM model on the tracker has no open alarm.
| Model | Coding | Reasoning | Tool use | Canary, hourly |
|---|---|---|---|---|
| GLM-5.3 | No declinedetector at 0.00 of 0.3024-hour score 89 | No declinedetector at 0.00 of 0.3024-hour score 94 | No declinedetector at 0.00 of 0.5024-hour score 88 | No alarmnone in 30 days |
03
Canary alarms in the last 30 days
Every hour the canary sends two fixed probes to every model and tests the last 6 and 24 hours against the prior seven days. An alarm opens at a 12-point fall with p < 0.01, and it closes once the 24-hour average is back within 10 points of the baseline.
Canary score fell from 79.4 over the prior seven days to 50.0 over the last 6 hours (Welch’s t-test, p = 0.008).
Closed after 1 h
Canary score fell from 96.0 over the prior seven days to 79.2 over the last 6 hours (Welch’s t-test, p = 0.005).
Closed after 2 h
Canary score fell from 98.0 over the prior seven days to 79.2 over the last 6 hours (Welch’s t-test, p = 0.003).
Closed after 3 h
04
How the tracker decides
What each state means. Every one comes from a published rule, and none is set by hand.
- Decline detected
- The suite’s Page-Hinkley detector fired on its newest reading: a sustained fall against the model’s own past, on daily medians. The constants are δ 0.01 and λ 0.30 for coding and reasoning; tool use needs λ 0.50 on three daily runs in a row. A detection resets the detector, which then re-learns the model’s new level over three daily runs, so the state lasts until the suite’s next run. While it shows, it is the alarm the Smart Router acts on.
- Confirming
- Tool use only. The statistic is over its threshold, and it has to stay there for three daily runs before it counts as a detection.
- Below baseline
- The leaderboard’s own early warning: the median of the newest runs is more than 5 points, and more than three standard errors, below the runs before them. It stays a watch, not an alarm, until the detector confirms it.
- No decline
- The detector is armed and its newest reading did not fire. The bar is its statistic against the threshold, so a full bar is an alarm. Right after a detection the detector is re-learning the model’s level, so a model that settled lower can read No decline here while the board still shows it below baseline.
- Calibrating
- A detector arms after ten days of daily results on the current version of the tests, so a new model, or a suite whose tests changed, starts here. A suite with no results yet on the current tests reads No baseline yet.
- Not checked
- The drift data did not answer when the page was built. It says nothing about the model, and the next refresh tries again.
- Canary alarm
- The hourly probe score fell 12 points or more against the prior seven days, with p < 0.01 on Welch’s t-test. It closes on its own once the 24-hour average is back within 10 points of that baseline.
The tolerance and thresholds were set against the real score history. How often the detector raises a false alarm is measured in How often our drift detector is wrong, measured. The full method is on AI drift detection, and every weight and constant is on the methodology page.
05
The Smart Router reads the same alarms
A tracker tells you a model got worse. The Smart Router acts on the same alarms, request by request.
When a model’s detector fires on a suite your routing mode ranks on, or the model has an open canary alarm, the Smart Router puts it behind every model without an alarm for new requests. Coding alarms count in the coding modes, reasoning alarms in the reasoning modes and tool-use alarms in the tool-use modes; agentic coding counts coding and tool use, the modes that rank on the combined score count all three, and a canary alarm counts in every mode.
The model is never removed. If every candidate has an alarm, the best of them still answers. A conversation the model is already answering stays on it, so an agent does not change models mid-task, and in a custom traffic split a flagged model gives way to the split’s other models first. The X-AISM-Reasoning header on each response names any model that was passed over and why.
The router acts for exactly as long as the alarm shows here: a detection until the suite’s next run, a canary alarm until it closes. It is on by default; the setting is “Avoid models with an active drift alert” in your router preferences. How each mode ranks is in the router documentation.
1,000 routed requests a month free · your own provider keys · no card
06
Questions about the degradation tracker
What is an AI degradation tracker?
A page that re-tests AI models on fixed tasks and tells you when one has got worse. This one covers the 25 models on the AI Stupid Level board, from OpenAI (GPT), Anthropic (Claude), Google (Gemini), DeepSeek, Moonshot AI (Kimi) and Z.ai (GLM), through each provider’s own API: coding every four hours, reasoning and tool use daily, with an hourly canary probe. Every suite has its own detector, run against the model’s own past, so a decline shows up as a dated measurement rather than a hunch.
How do you know a model got worse, and not just had a bad day?
One bad run proves nothing: model output varies from run to run even when nothing has changed upstream. Each suite’s Page-Hinkley detector adds up how far every day’s median falls below the model’s running mean, less a small tolerance, so noise cancels out and only a sustained fall reaches the threshold. Sharp drops are the hourly canary’s job, and it needs a 12-point fall with p < 0.01 before it opens an alarm. How often the detector is wrong is measured in How often our drift detector is wrong, measured.
Is Claude, GPT or Gemini getting worse right now?
As of 30 Oct 2026, 16:16 UTC: None of the 9 GPT models on the tracker has an open alarm. None of the 9 Claude models on the tracker has an open alarm. None of the 3 Gemini models on the tracker has an open alarm. None of the 2 DeepSeek models on the tracker has an open alarm. The Kimi model on the tracker has no open alarm. The GLM model on the tracker has no open alarm. Each family also has its own page with the 24-hour scores and the dated record: is ChatGPT getting worse, is Claude getting worse, is Gemini getting worse and is DeepSeek getting worse.
Does the Smart Router avoid degraded models?
Yes, by default, for new requests. A model whose detector has fired on a suite your routing mode ranks on, or that has an open canary alarm, goes behind every model without an alarm. It is never removed: if every option has an alarm the best of them still answers, and a conversation that model is already answering stays on it. The X-AISM-Reasoning header on each response names any model that was passed over. The setting is “Avoid models with an active drift alert” in the Smart Router preferences.
Can I get an email when a model I use degrades?
Yes. A free account includes a watchlist of three models with an email when a watched model’s coding score falls five points below a week earlier, and a weekly summary of what moved.
How often is each model tested?
Coding every four hours, seven tasks with seven attempts each; reasoning and tool use once a day; and the canary’s two fixed probes every hour. The detectors run on daily medians, so they confirm a change over days, while the canary is the fast instrument for a collapse. For five models, reasoning and tool use are funded by the community and run when a member funds a run; their cells say so.
Why does a new model show Calibrating?
A detector arms after ten days of daily results on the current version of the tests. Until then there is not enough history to tell a change from the model’s normal spread, so the tracker says so instead of guessing. A suite whose tests were changed starts again the same way.
Is the degradation tracker free?
Yes, the tracker is public, like the model pages, the methodology and the status page. Free account: every ranking over 24 hours, 7 days and 30 days, a 3-model watchlist with email alerts, 1,000 routed requests a month and a Data API key. No card.
Know the next time a model you use degrades
Free account: every ranking over 24 hours, 7 days and 30 days, a 3-model watchlist with email alerts, 1,000 routed requests a month and a Data API key. No card.
Three models free · an email when a watched model’s coding score falls five points below a week earlier, and a weekly summary of what moved