One provider is
one point of failure.
On 3 September 2026 three frontier assistants went down on the same day. Routing across many upstreams is not an optimisation. It is the difference between a bad night for a provider and a bad night for your product.
// Why this matters
Four months of somebody else’s bad nights.
Every figure below is published by somebody other than us. Independent trackers count incidents differently, so we cite the outage rather than argue about the tally.
47
Provider incidents, June to September 2026
Logged on one major provider's own status page across four months, twenty-seven of them in July alone.
status.openai.com, retrieved September 2026
3h 6m
Longest single outage that quarter
On 3 September 2026 ChatGPT, Claude and Grok were all down on the same day; one of them for three hours and six minutes.
The Register, 3 September 2026
39×
Speed spread on one model
Latency for the identical model differs by a factor of thirty-nine depending on which route serves it.
Artificial Analysis provider comparison, September 2026
8.9×
Price spread on one model
Across eighteen providers, the same model costs up to nine times more depending on where you buy it, and the spread moves weekly.
Artificial Analysis provider comparison, September 2026
// Four things that never happen
What failover will not do to you.
- 01
Never a different model
Failover only moves a request to another endpoint serving the same published model id. If no endpoint can serve it under your policy, the call fails. We do not quietly substitute something cheaper and hope you do not notice.
- 02
Never a double charge
Every generation emits exactly one idempotent usage event, whatever happened upstream. A retried request that produced one answer bills once.
- 03
Never past the ceiling
Budget is reserved before dispatch and settled afterwards against shared hot state. A key at its limit stops at the limit, not at the end of the month.
- 04
Never out of region
A key pinned to EU routes fails closed when no in-region endpoint can serve it. Residency is not a best-effort preference.
// Engineering targets
The numbers we hold ourselves to.
These are targets on the part of the path we control, not measurements of the models behind it, and not vendor benchmarks. We would rather publish the target and be held to it than quote somebody else’s number.
Pre-dispatch overhead
p50 ≤ 10 ms · p95 ≤ 25 ms
Auth, policy, budget and catalog resolve: the part of the path we own.
Chunk-forward lag
p95 ≤ 5 ms
Added latency between receiving an upstream chunk and forwarding it to you.
New provider integration
≤ 1 day
An OpenAI-compatible upstream is one manifest file and no new code.
Retention purge proof
Hourly
Every artifact carries a data class and a retention class, verified on a schedule.
Always on, under every policy.
Automatic retry and failover, per-key budgets and spend alerts run underneath all five routing policies. Custom policies are available on committed-use accounts.
Cheapest
The lowest price that still meets your model and feature needs.
Fastest
The lowest latency right now, for live products and agents.
EU-only
Inference stays on EU routes, enforced per key.
Global
Routes across every available provider, for the widest choice.
Quality-first
The strongest model for the task, with a fallback ready.