Reproducible model aliases
Hard-coding model names ages badly. Zumik aliases let you ask for fast coding or best reasoning, then resolve that through an immutable release that records exactly which model answered and why, so routing stays auditable and replayable.
Stable names for the decisions you actually make.
auto.bestHighest-quality response under current policy.
auto.balancedCost-quality tradeoff with caching enabled.
auto.fastLowest-latency frontier response.
auto.cheapestCost-minimized path, often open-weights.
code.fastLow-latency coding tasks.
code.balancedQuality coding with prompt caching.
reasoning.bestDeep reasoning and planning.
vision.balancedMultimodal tasks.
batch.standardNon-interactive background work on batch tiers.
Frozen policy, versioned forever.
An alias release captures the resolution policy at a point in time. The live alias points at one release; a new provider-model revision produces a new release. Your logs carry the release id, so any decision is explainable months later.
See the routing mix{
"requested_model": "code.fast",
"alias_release": "alr_2026_06_09_003",
"resolved_model": "anthropic/claude-haiku-4-5",
"candidates_considered": 3,
"resolution_reason": "lowest_expected_latency_under_policy"
}The contract that keeps aliases honest.
Alias releases are immutable.
A live alias points to exactly one release at a time.
Replays pin a release or a concrete resolved target.
Customer logs expose the alias release id.
/v1/models returns exact upstream-compatible objects.
A provider-model revision change creates a new release, frozen aliases never change silently.
Aliases give you the convenience of automatic routing without the usual cost: opacity. You get the router’s decision and the receipt for it.
Each alias maps to real models in the catalog.
Every model page lists the aliases it serves. Compare the candidates an alias chooses between, and see why one wins under current policy.
Model aliases, answered.
What is a model alias?
A stable logical name like code.fast that resolves through an immutable release to a concrete provider model, so applications request a capability rather than hard-coding a model string.
Why are alias releases immutable?
So routing is reproducible. A live alias points at one immutable release; changing a provider-model revision creates a new release rather than mutating the old one. Past decisions can always be explained.
Does /v1/models still work?
Yes. /v1/models returns exact upstream-compatible model objects. Rich alias metadata lives under /v2/model-aliases so the OpenAI surface stays clean.
Can a replay reproduce an old routing decision?
Yes. A replay pins an alias release or concrete resolved target, so it gets the same decision later even after the live alias has moved on.
Route by intent. Audit by default.
Use aliases like code.fast and reasoning.best, and get a reproducible record of every decision.