Is DeepSeek getting worse?

Places are the leaderboard's own: walking down a board, a model joins the group above it unless that group's leader is ahead by more than the measurement can resolve, so “joint” places are normal. A detector is armed once it has ten days of daily medians on the current test version; it fires only on a sustained fall, never on one bad run. Every DeepSeek model has its own page with its history, and all of them are on DeepSeek models.

Down is availability: every 10 minutes we send one short completion to api.deepseek.com on our own key and record whether it answered, how fast, and whose fault a failure was — the provider's, our account's, the probe's or our server's. Downgraded is quality: the scores and detectors above, from fixed tasks the API did answer. A model can be fully available and quietly worse, or briefly unavailable and unchanged.

We measure the API (api.deepseek.com), not the consumer app. The DeepSeek app is not the DeepSeek API: we probe the API, which is what applications and the Smart Router call.

The 7-day history and every episode are on DeepSeek API status; all six providers are on the status hub. When a provider stops answering, the Smart Router's fallback order moves traffic to the next model on the board.

The DeepSeek app and the DeepSeek API are two surfaces of the same models. The app adds its own prompt, tools and product logic; the API at api.deepseek.com is what applications call, and it is what AI Stupid Level tests — model by model, with fixed tasks, the same prompts and the same output budget every time, with nothing of yours in between.

That makes the scores a control. If the API-side scores hold and the detectors stay quiet while the DeepSeek app feels worse, the change is on the tool side: the harness, the context it assembles, or a session that has grown long enough to hurt (see context rot). If the model behind the name changes, it shows up here, on the same tasks, inside the detector's confirmation window.

The whole record, every vendor, is on Is your AI getting worse?; the method and its measured false-alarm rate are in How often our drift detector is wrong, measured.

Page-Hinkley change-point detection (a CUSUM-family test) run per suite on each model’s daily median, constants published: δ 0.01; λ 0.30 for coding and reasoning, λ 0.50 held for three consecutive daily runs for tool use; detectors run on daily medians and are armed after ten days on the current test version, so a real shift is confirmed over days, not hours. An hourly canary of two fixed probes catches sharp drops with Welch’s t-test (12-point fall, p < 0.01).

Every score carries a standard error from its own run-to-run repeatability; models whose lead is inside that noise share a place, and a composite is labelled with how many suites contributed. A prompt a provider declines is recorded as a refusal with the provider's stated category and that session's score is withheld, never imputed, so a tightened filter shows as a refusal rate rather than as invented incapability.

The detector, its constants and its measured false-alarm rate are on AI drift detection and in the methodology. The methodology is public — every weight, statistical test and drift constant. The test tasks are private, so no lab can train on them.

Watch the DeepSeek models you depend on and let the boards do the checking: an email when a watched model’s coding score falls five points below a week earlier, and a weekly summary of what moved. If you build on the API, the Smart Router sends each request to the model that scores best for it across every vendor on the board and is drift-aware by default, on your own provider keys.

The same question for the other families: Is ChatGPT getting worse?, Is Claude getting worse? and Is Gemini getting worse?. Every model on the board, with the whole record, is on Is your AI getting worse?; the DeepSeek models' scores, prices and head-to-head pages are on DeepSeek models.