Skip to main content
Forge recommends the cheapest capable model for a task before dispatch, from a deterministic rubric you can read in the repo (src/model_tiers.json). Unlike a gateway that decides inside the proxy at request time, the routing decision is visible and diffable in git.

​
Recommend a tier — forge route

The recommendation is exemplar k-NN math over a labeled bank (English + Hinglish rows) under an overlap-similarity metric with a confidence gate — not a keyword lookup.
Start at the recommended route.tier and escalate only after an external verifier fails, never preemptively. This keeps spend down without capping capability when a task genuinely needs it.

​
Intent, then tier

Routing shares its math with intent detection (src/intent.js): a prompt maps to an intent by the same exemplar k-NN estimator. Note the two use different stop-sets — route treats generic verbs (fix / add / build) as complexity noise, but those verbs are exactly the intent signal.

​
The tier table

The tier table (src/model_tiers.json) pins public Anthropic model IDs by family (haiku / sonnet / opus / fable). Doc prices are reconciled against this file by the docs check, so prose and the table can’t drift.

​
Self-hosted gateway remap

A self-hosted LiteLLM or proxy gateway serves its own model names, so a stock ID sent verbatim would 404. When a non-default gateway base URL is configured, Forge (src/gateway_model_map.js) fetches GET /v1/models once per process and scores each advertised id against every tier’s family:
1

Hard family gate

The family word (haiku / sonnet / opus / fable) must match — it is a hard gate.
2

Best overlap wins

Among the family, the setOverlap coefficient of the tier’s name tokens picks the best match.
3

Tie-break to canonical

Ties break toward the id closest to the canonical name.
The remap consults the gateway only when the resolved id is a stock ID — an explicit .forge/providers.json alias or an ANTHROPIC_MODEL override is never touched. It fails safe to the stock ID on no gateway, an unreachable /v1/models, or no family match, so direct api.anthropic.com users are byte-identical.
forge doctor’s gateway models row prints the resolved tier → model mapping for verification.

​
Providers and cost

forge cost --stages reports measured stages only — a stage with no events says “no data”, never a default. A number is an assumption until measured.