System status

Real numbers from this site and its AI assistant, aggregated every hour. No question text and nothing that identifies a visitor is ever published.

Checking statusLive status needs JavaScript

Loading the latest numbers.This page was built without live numbers. With JavaScript they load and refresh every minute.

Page built 25 Oct 2026, 21:13 UTC

Availability

Last 30 days, one synthetic check an hour

LoadingNot in this snapshot

26 Sep: no checks27 Sep: no checks28 Sep: no checks29 Sep: no checks30 Sep: no checks1 Oct: no checks2 Oct: no checks3 Oct: no checks4 Oct: no checks5 Oct: no checks6 Oct: no checks7 Oct: no checks8 Oct: no checks9 Oct: no checks10 Oct: no checks11 Oct: no checks12 Oct: no checks13 Oct: no checks14 Oct: no checks15 Oct: no checks16 Oct: no checks17 Oct: no checks18 Oct: no checks19 Oct: no checks20 Oct: no checks21 Oct: no checks22 Oct: no checks23 Oct: no checks24 Oct: no checks25 Oct: no checks

Every hour the Worker checks its own health endpoint.

Conversations

Per day, last 30 days

LoadingNot in this snapshot

26 Sep: not measured27 Sep: not measured28 Sep: not measured29 Sep: not measured30 Sep: not measured1 Oct: not measured2 Oct: not measured3 Oct: not measured4 Oct: not measured5 Oct: not measured6 Oct: not measured7 Oct: not measured8 Oct: not measured9 Oct: not measured10 Oct: not measured11 Oct: not measured12 Oct: not measured13 Oct: not measured14 Oct: not measured15 Oct: not measured16 Oct: not measured17 Oct: not measured18 Oct: not measured19 Oct: not measured20 Oct: not measured21 Oct: not measured22 Oct: not measured23 Oct: not measured24 Oct: not measured25 Oct: not measured

Counted from the anonymous question log, hour by hour.

Time to first token

From question received to the first word of the answer

LoadingNot in this snapshot

Percentiles need at least 5 conversations to mean anything.

Cost

Tokens reported per answer, times OpenAI list prices

LoadingNot in this snapshot

Model and embedding tokens, priced per model.

Prompt cache

Input tokens served from OpenAI's prompt cache

LoadingNot in this snapshot

Cached tokens cost a tenth of fresh ones.

Failed answers

Model unavailable, answer interrupted or server error

LoadingNot in this snapshot

A question that fails before the first word is given back to the visitor.

Checked answers

New assistant: each sentence checked against his published facts

LoadingNot in this snapshot

Answers that fail a check are cut or replaced, never shown.

Interest

Visitors who used all 10 questions, then got in touch

LoadingNot in this snapshot

Only the click on Email or LinkedIn is recorded, once per visitor.

Evaluation

Golden set of 60 questions, last reviewed run

recall@5
95.7%
time to first token, p95
1,525 ms
hallucination bait rejected
100%

Last run of the pre-launch review, 8 Oct 2026, with the real model on the preview deployment, as published on How it works.

How this is measured

Status and availability
Every hour the Worker calls its own health endpoint and makes one database round trip, and records whether both answered within 5 seconds. Degraded means the latest check failed, or more than 1 in 5 questions failed in the last 24 hours. Paused means the assistant's switch is off while the site stays up.
Conversations
A conversation is a run of questions from the same anonymous visitor with less than 30 minutes between them. The source is a question log that stores no IP address or browser details and is deleted after 90 days. Only hourly counts are published, never a question.
Time to first token
Measured in the Worker, from receiving a question to sending the first word of the answer, so retrieval is included and the network to your browser is not. p50 and p95 are exact percentiles over every answer in the window, published once there are 5 conversations.
Cost
The tokens OpenAI reports for each answer (input, cached input, cache writes and output) times the list price of the model that answered, plus an estimate for the query embedding. Prices were last checked on 7 Oct 2026 against OpenAI's pricing page.
Prompt cache
The share of input tokens OpenAI served from its prompt cache. The fixed instructions and profile come first in every prompt, so OpenAI can reuse them across questions at a tenth of the input price.
Interest
Of the visitors who asked anything in the last 30 days, the share who used all 10 questions, and of those, the share who then clicked Email or LinkedIn on the limit card. Only that click is recorded, once per visitor.
Evaluation
Scores from the latest run of the golden question set, bundled when this page was built: how faithful the answers are to the sources, whether retrieval finds the right ones, and whether any answer leaked the instructions. When no run was bundled, the golden-set figures of the pre-launch review, the ones How it works publishes, with the day they were taken.

Prices from OpenAI's pricing page, standard tier. How visitor data is handled: privacy notice.