BLOCK 4 · AI GATEWAY (PRODUCTION) 3:10 – 3:20 · 10 min · 3L / 7HO · Lab

TOPIC 4.3 / 10 · HO #2

Configure models + aliases

Three concepts in the catalog: available, enabled, aliased. Aliases let production teams swap models without touching app code.

4.3

Models + aliases

10 min 3L / 7HO 3:10 – 3:20

Slide 1 / 3 · The catalog

Which models are available, default, aliased

Available — what your providers expose
Enabled — what your gateway accepts requests for
Aliased — internal name → real provider/model id
Aliases are how production teams swap models without code changes
Two levers, two layers: aliases swap a model id; routing rules (coming in 4.6) swap a provider. Same goal — change behavior without touching the app.

Slide 2 / 3 · The alias trick

Why aliases matter

Before (no alias):

model: "openai/gpt-4o-mini"

After (alias):

model: "small-fast"

alias: "small-fast" → "openai/gpt-4o-mini"

Model swap = config change. Not deploy.

Slide 3 / 3 · Five clicks

HO #2 — Configure models (7 min)

Sidebar → Models → Model Catalog
Enable Claude Sonnet 4.6 (Anthropic)
Enable gpt-4o-mini (OpenAI)
(Optional) Models → Model Providers → OpenAI → Edit key → Aliases (Optional) → Name small-fast · Value gpt-4o-mini
Confirm both models show as enabled

Per-provider note: aliases live on the provider key, not gateway-wide. The mapping you add to OpenAI's key only resolves when the request is routed to OpenAI. Edit other providers' keys to alias the same name across providers, or use a routing rule + alias combo.

Speaker notes

Slide 1: "Three concepts in the catalog. Available is what your providers expose — could be 30 models. Enabled is what your gateway accepts requests for — should be a curated subset. Aliased is an internal name that points to a real model id. Aliases are how production teams swap models without changing code."

Slide 2 — the alias trick: "Imagine you have 47 places in your code that say model: openai/gpt-4o-mini. OpenAI deprecates that model. You're doing a 47-file search-and-replace. With an alias: your code says model: small-fast. The gateway config says small-fast → openai/gpt-4o-mini. When OpenAI deprecates, you update the gateway. One change. Production stays up. (Note the provider/model format — Bifrost's chat-completions endpoint expects it. Aliases let you hide that prefix from app code, which is half the point.)"

Slide 3 — HO: "Seven minutes. Model Catalog. Enable one Anthropic model — Claude Sonnet 4.6. Enable one OpenAI model — gpt-4o-mini. Optionally add the small-fast alias — and here's where Bifrost surprised me: aliases don't live at the gateway level, they live on the provider key itself. So you go to Model Providers, click into OpenAI, hit Edit key, and you'll see an Aliases section at the bottom of that form. Add small-fast as the Name, gpt-4o-mini as the Value. That alias only resolves when traffic is routed to OpenAI — which makes sense once you think about it: the alias is mapping a custom name to OpenAI's specific deployment/profile id. Loop the room. While you're in that form, also note the Allowed Models and Blocked Models fields — that's a per-provider-key allowlist, distinct from the VK-level allowlist we'll see in 4.4. And the info banner at the bottom — 'synced from config.json' — that's the precedence rule from 4.5.5 spelled out at the point of edit."

Triage note: Tier 1 cut drops the alias step (saves ~2 min). Alias concept lands in the lecture; the live add is optional. If room is on time, do the alias — it's memorable.

Demo piece — Attendees

Click sequence:

Sidebar → Models → Model Catalog
Find Claude Sonnet 4.6 → toggle Enable
Find gpt-4o-mini → toggle Enable
(Optional) Sidebar → Models → Model Providers → OpenAI → Edit key. Scroll to the Aliases (Optional) section. In the Name column type small-fast; in the Value column type gpt-4o-mini (the provider's identifier; can also use deployment names, inference profile IDs, fine-tuned endpoint IDs). Save.

Per-provider note: Bifrost shows you the alias help text: "Map each request model name to the provider's identifier (deployment name, inference profile ID, fine-tuned endpoint ID, etc.) or just a custom name, e.g. claude-sonnet-4-5 → custom-claude-4.5-sonnet." The alias is per provider key — if you want small-fast to also resolve when Bifrost routes to Anthropic, you'd add the alias on Anthropic's key too (mapped to claude-haiku-4-5 or similar).

Sync banner to point at on the form: "This config is synced from config.json. Any future updates to config.json will overwrite UI changes." That's the exact precedence rule from Beat 4.5.5 — Bifrost makes it visible at the point of edit.

Instructor mirror on projector.