# modelstatus.dev

> Availability, latency and throughput history for LLM endpoints, reconstructed
> from OpenRouter's rolling-window metrics. 400 models and
> 1200 provider endpoints are sampled every 5 minutes. All times UTC.

OpenRouter publishes current uptime, latency and throughput per model endpoint,
but no history: there is no way to ask what a provider looked like an hour ago.
This site samples those metrics every five minutes and keeps the series.

Two caveats that matter when quoting these numbers:

- Latency and throughput are OpenRouter's pre-aggregated 30-minute rolling
  windows, not raw measurements. Hourly figures are averages of those windows, so
  an hourly "p99" is the mean of twelve p99 readings, not a true hourly p99.
  Quantiles cannot be averaged.
- Uptime over 5 minutes is exact, because consecutive 5-minute windows do not
  overlap at a 5-minute sampling interval.

## Live status

- [Current status](https://modelstatus.dev/llms-full.txt): every endpoint currently down or
  degraded, incidents in progress, and per-provider counts, as plain text.

## Pages

- [Status dashboard](https://modelstatus.dev/): incidents in progress, recently resolved, and
  endpoints that have been down for more than six hours.
- [All models](https://modelstatus.dev/models): the full catalogue grouped by author.
- [Incident archive](https://modelstatus.dev/incidents): every recorded incident, filterable by
  provider and severity, plus an [RSS feed](https://modelstatus.dev/incidents.xml). Each model
  also has its own feed at `https://modelstatus.dev/models/{author}/{slug}/incidents.xml`.
- [Changes](https://modelstatus.dev/changes): new and retired models, pricing transitions and
  incidents over the last 7/14/30 days.
- [API & feeds](https://modelstatus.dev/api): every machine-readable surface — the status JSON
  endpoint, RSS feeds and status badges — documented for human readers.
- [Collection health](https://modelstatus.dev/health): how the 5-minute sampling pipeline itself
  is doing.
- Status JSON: `https://modelstatus.dev/api/v1/status.json` — the current incident/totals
  document, refreshed every 5 minutes.
- Status badges: `https://modelstatus.dev/badge/{author}/{slug}.svg` and
  `https://modelstatus.dev/badge/provider/{provider}.svg`.
- Compare models: `https://modelstatus.dev/compare/{author}/{slug}/{author}/{slug}` — up to four
  models on shared axes, each shown as the best value across its providers.
  Example: [openai/gpt-oss-120b/openai/gpt-oss-20b](https://modelstatus.dev/compare/openai/gpt-oss-120b/openai/gpt-oss-20b)
- Compare a model's providers: `https://modelstatus.dev/compare/endpoints/{author}/{slug}` —
  every endpoint of one model side by side, with per-column best values and a
  price-vs-latency scatter. Example: [openai/gpt-3.5-turbo](https://modelstatus.dev/compare/endpoints/openai/gpt-3.5-turbo)
- Model pages carry a half-year daily availability calendar and the full
  pricing-transition journal per endpoint (including cache-read/write prices).
- [Sitemap](https://modelstatus.dev/sitemap.xml): every model page.

## Models by author

- [openai](https://modelstatus.dev/models): 94 models
- [qwen](https://modelstatus.dev/models): 51 models
- [google](https://modelstatus.dev/models): 41 models
- [anthropic](https://modelstatus.dev/models): 28 models
- [mistralai](https://modelstatus.dev/models): 19 models
- [z-ai](https://modelstatus.dev/models): 15 models
- [deepseek](https://modelstatus.dev/models): 13 models
- [nvidia](https://modelstatus.dev/models): 13 models
- [minimax](https://modelstatus.dev/models): 9 models
- [meta-llama](https://modelstatus.dev/models): 8 models
- [moonshotai](https://modelstatus.dev/models): 8 models
- [bytedance-seed](https://modelstatus.dev/models): 6 models

## Optional

- [Source code](https://github.com/Panmax/model-status): collection pipeline and
  schema, including why a missed round cannot be backfilled.