How Clauder predicts Claude Code limits
The math behind the 5-hour ETA, the weekly projection, and the per-account survival model.
Clauder predicts your 5-hour limit by dividing the percentage you have left by your current burn rate. It projects your weekly limit by replaying your hour-by-hour pattern from the last four weeks. Both run on clauder.fyi from the percentages your extension syncs.
This page walks through the 5-hour ETA, how confidence decides what you see, the weekly projection, and how scored predictions train a model for your account.
How does the 5-hour limit prediction work?
It divides the percentage left in your 5-hour window by how fast you are using it. At 60% with a burn rate of 0.5% per minute, the ETA is 40 / 0.5 = 80 minutes.
- Samples. Each sync reports your current 5-hour percentage. The server turns changes into burn-rate samples at least 30 seconds apart and keeps the latest 10.
- Smoothing. The burn rate is an exponential moving average with weight 0.3 on the newest sample, raised to 0.5 when you come back from a break of 5 minutes or more so the rate catches up quickly.
- Range. One standard deviation of the rate above and below gives an earliest and latest ETA. The range is shown when the two are at least 5 minutes apart.
- Idle sessions. With no new usage for 7 minutes, the session counts as idle. New samples stop and confidence drops.
- Window resets. A drop of more than 30 points after 30 minutes idle, or a fall from above 50% to below 10%, marks a new window and clears the samples.
How is prediction confidence scored?
Every prediction gets a score from 0 to 1. The score sets how precise the ETA you see is, and whether you see one at all.
The 5-hour score starts at 0.4. It rises with 3 or more samples and with an active session. It falls when the burn rate swings, when you are idle, and when the ETA is more than 2 hours away.
| Tier | Score | What you see |
|---|---|---|
| High | 0.75 and up | A precise ETA, such as ~25 minutes or ~1h 20m |
| Medium | 0.50 to 0.74 | A precise ETA |
| Low | 0.35 to 0.49 | A rounded bucket, such as ~30-45 min or ~2+ hours |
| Insufficient | Below 0.35 | No ETA |
How does the weekly limit prediction work?
It replays your usual week. The server averages your token use for each of the 168 hours of the week over the last 28 days, in your timezone, and adds those hours to your current weekly percentage until the week ends.
- Tokens to percent. Anthropic's weekly limit is not a token count, so the server calibrates tokens against the weekly percentage you have used so far.
- Trend. A trend factor compares your recent days with your history and scales the remaining hours up or down.
- Breach day. The day and hour you would hit the limit appear only when the projection passes 110% and confidence is medium or high.
- Confidence. The weekly score grows with 7 and then 14 days of data and with a consistent day-of-week pattern. It is capped at 0.85.
How do predictions get more accurate over time?
Clauder checks its own predictions against what happened, and those scored predictions are the training data for a survival model fitted to your account.
- Snapshots. Predictions with at least low confidence are saved: at most one 5-hour prediction every 5 minutes and one weekly prediction per hour.
- Scoring. When you hit the 5-hour limit, each open prediction from the previous 5 hours records its error in minutes. When the window resets without a hit, open predictions close as misses.
- Survival model. Training a Cox proportional hazards model takes at least 20 scored predictions that include 3 or more limit hits. Its inputs are your 5-hour percentage, burn rate, hour of day, day of week, and the confidence score.
- Model ETA. Once your account has a trained model, the 5-hour ETA is the model's median time to the limit. It also gives the times by which you have a 70% and a 90% chance of hitting it. Confidence then comes from the model's concordance index: 0.8 and up is high, 0.65 medium, 0.5 low.
Without a trained model, the burn-rate ETA above is used. The statusline marks model-based ETAs with a lightning icon.
Why don't I see a prediction?
- Sync is off. Predictions need
clauder.sync.enabledand a license key from an active Pro plan or trial. See getting started. - Confidence is insufficient, usually right after a reset or with few samples.
- The session is idle, which hides the burn-rate slide in the status bar.
- The last prediction is more than 5 minutes old, for example after losing the network. The extension hides it rather than show a stale ETA.
The server caches each prediction for 30 seconds. The data and privacy page lists the fields a sync sends.