Measured, not promised.

Numbers from our own instrument, with date, method and limits. When independent comparisons exist, we will publish them here too.

First-response latency

From your question to the first words of the answer, with a real browser against the product. Bars scaled to 6 seconds.

Normal conversation (steady state)

0.3 to 1.6 s

From your question to the first words of the answer.

Long input (3,394 characters)

1.9 s

A very long message, with all its context.

Thinking mode (complex problem)

5.8 s

When the work requires reasoning in steps before answering.

71 real-use scenarios

Full functional marathon with zero freezes and zero page errors, including long tasks and documents.

3 concurrent sessions

Simultaneous use from several devices with no noticeable degradation.

Per-user isolation

Container-escape and prompt-injection tests passed: every person lives in their own space.

Documents and attachments

A work document retrieved and analyzed in about 5 seconds in fast mode; attachments purged after 24 hours.

Vision

Accurate image descriptions in 4 to 8 seconds after the image mode fix.

No telemetry

No measurement leaves your account to third parties: we measure, for you.

These measurements are ours: internal instrument, same conditions and dated. They are not an independent study.
Latency depends on the model provider and on your network; at peak hours it can go up. Our goal is to keep it in conversation ranges, not patience ranges.
We do not publish comparisons with other services using data we cannot verify. When serious independent tests exist, they will be published as they are, whether they favor us or not.

Want to reproduce these tests or propose others? Write to us on theDiscord or throughContact.