Status

The five slot-relay figures cover the last seven UTC calendar days, including today. They can include both Messages and token-count calls. One user request can produce more than one attempt when a safe fallback is tried. Requests rejected before reaching a slot relay are not counted.

Transfer completion means a successful HTTP reply (2xx) was fully transferred. Error replies, cancellations, and interrupted transfers are not successful. It does not tell us whether the model completed your task or guarantee service availability. The availability charts below measure something different.

The first-byte timer starts when the slot accepts the relay request and stops when the first upstream response body byte arrives. That byte may be stream metadata, not a generated word. It excludes the trip from your computer to the gateway and SSH setup. Full transfer time also grows with answer length. These are sample-weighted averages, not medians or percentiles; each timing can have a different sample count.

• Average upstream connection: 33 ms · 7,276 timing samples
• Average first response body byte: 3.32 s · 7,272 timing samples
• Average full transfer: 14.78 s · 7,302 timing samples

Traffic figures require at least 20 attempts. Each timing separately requires at least 20 samples, so a timing may remain unknown even when traffic totals can be shown.

Streamed output rate combines native/bridge slot-relay measurements and HTTPS gateway measurements without counting a gateway reply twice. It covers successful, fully delivered /v1/messages streaming replies. It divides the sum of provider-reported output tokens by the sum of time from the first to the last successfully forwarded response-body chunk at the observing relay. A reply needs valid output usage, at least two chunks and 100 ms of delivery time. Single-chunk or too-short replies are excluded; buffering can affect the observed rate. It excludes the initial wait, but includes stream pauses and network delivery, and includes all output tokens reported by the provider, including thinking or tool output where counted. Higher is faster for comparable models and workloads; this is an observed delivery rate, not a pure model-speed benchmark.

Output rate needs at least 20 eligible streamed replies over the same seven UTC calendar days. It can be shown while slot-relay figures remain unknown, or remain unknown when they are ready. Collection starts with new observations; older traffic is not backfilled. The gateway retains anonymous daily numeric totals; native relays keep account-bound daily numeric totals in the slot report. This output-rate panel publishes no user, slot or model breakdown or individual request timestamps. The homepage separately shows masked user usage and model details.

Only combined numbers are public, even if only one account contributes. This panel shows no account names, slots, messages, or individual request records.

Think of the gateway as a junction that passes requests along. These numbers show how busy that junction is. High CPU use, little available memory, long I/O waits or a nearly full disk can help explain slow replies. They do not prove that the gateway is the bottleneck: your network, the slot machine and Anthropic can also affect response time.

• Logical CPUs: 4 · one-minute load average: 0.00
• Host I/O wait: 0.0% · VPS CPU steal: 0.0%
• Data filesystem space available: 61.6 GiB

Host CPU counts time executing work; I/O wait and CPU time taken by the VPS provider are shown separately. Load average includes tasks running or waiting for I/O, so it is not a CPU percentage. Host memory used is total minus available memory. The storage card measures the filesystem that holds Fleet data, including other files on that filesystem; it does not measure all disks or disk throughput.

Fleet CPU use is a percentage of the service's configured CPU limit, rather than a percentage of the whole VPS. Fleet memory is container usage counted toward the limit, including cache and kernel memory; it differs from estimated host memory. Samples update about every 15 seconds; this page refreshes every minute. Missing or stale values remain unknown. Machine resource totals are independent of transfer samples.