The median agent step on Ostrid, from request to answer.
Ostrid runs AI agents on warm workers in three regions. Every run is measured, every step is on a timeline, and you can see where the time went.
One run, four moments.
A request arrives, the agent plans, tools answer, the model replies. Each frame below is one moment of the same run, left to right.
T + 0 MS
Receive
Request lands on the closest warm worker.
T + 120 MS
Plan
The agent picks its next tool.
T + 190 MS
Call tools
Search and code run, results stream back.
T + 310–340 MS
Respond
First token reaches your user.
Response time, last 30 days.
2.4M runs · all regions
Latest releases
2.14
Warm pools for long-running agents
Agents that run longer than a minute now keep their worker warm between steps. Tool calls no longer pay a cold start.
340 ms
2.13
Streaming tool results
Tool output streams back to the model as it arrives, so a slow search no longer blocks the next token.
410 ms
2.12
Run replay in the console
Every run can be replayed step by step, with the exact inputs, timings and tool responses it saw.
520 ms
“We sent the first query at 09:14 and were live by 09:31. The run view showed us which tool was slow before we asked.”
Ama Lindqvist-Osei, Head of Platform, Tallowfinch
09:14
09:31
“Our p99 used to be a rumour. Now it is a number on the wall, and it keeps going down.”
Rhys Okafor-Bell, CTO, Tidewick
Pricing
Starter
$0
Shared queue. Workers go cold after 60 s idle.
Scale
$480 / mo
Warm pools and queue priority: 40 ms less wait at peak.
Dedicated
Custom
Reserved workers in your region. p99 under 900 ms.
Questions
What counts as a run?
One request to an agent, from the first token in to the last token out, including every tool call in between.
Where do runs execute?
In Lisbon, Frankfurt and Toronto. Requests go to the closest healthy region unless your key is pinned to one.
How do you measure latency?
From the moment a request reaches our edge to the moment the first token leaves it. We publish p50, p95 and p99, never averages.
Can I bring my own model?
Yes. Dedicated plans can run your weights on reserved workers. Scale plans can call any model with an API.
What happens if a region goes down?
Traffic moves to the next closest region within one health check. Pinned keys fail over only if you allow it.
now