≥85%
Stays in the shallows
Target share of unpinned steps served by Tier-0 and Tier-1 models. Spend falls through routing, not through a thinner margin.
AI gateway · prepaid · OpenAI-compatible
Curiosity traffic does not need frontier models, and builders rarely do either. idlidu.dev routes to the shallow end by default, charges you a named markup on what actually ran, and stops dead at your prepaid cap. Pin a frontier model any time you mean to pay for one.
A worked example, not a benchmark — the rates come straight from the catalogue and the markup from the billing constant.
≥85%
Target share of unpinned steps served by Tier-0 and Tier-1 models. Spend falls through routing, not through a thinner margin.
1.5×
Provider list price times a number we print. No silent usage padding, no inflated token counts, no opaque pool.
$0
Prepaid only. The gateway refuses the call that would exceed your balance or a key's hard limit, including across parallel trees.
It is the OpenAI API. Change the base URL, keep your SDK, your types and your
retries. model: "bottom" picks the cheapest capable model;
any catalogue id pins it instead.
from openai import OpenAI
client = OpenAI(
base_url="https://gateway.idlidu.dev/v1",
api_key="idl_...",
)
r = client.chat.completions.create(
model="bottom", # cheapest capable model, or pin an id
messages=[{"role": "user", "content": "Summarise this diff"}],
)
print(r.usage) # what ran, and what it costidd · ego · sego
Drive, reality and limits, after Freud's three-part model of the mind. One catalogue underneath all three, and three different answers to the question every other tool settles quietly in a system prompt you never see: who decides what you are allowed to want?
Drive. No policy layer, by design.
Reality. Proposes, then proves. Launches here.
Limits. Guardian-set, for under-18s.