pi-cheaperinference

A pi package that registers CheaperInference as a first-class provider: model selection from the live catalog with native pricing, so pi's usage/cost reporting is correct out of the box — no manual provider or cost configuration.

Install

pi install npm:@lilco/pi-cheaperinference

Setup

export CHEAPERINFERENCE_API_KEY=ci-...

Get a key at cheaperinference.com (referral link — it supports this plugin at no cost to you).

Then pick a cheaperinference/* model in pi (/model or pi --model cheaperinference/gpt-5.6-luna).

What it does

• Fetches the authenticated catalog (GET /v1/models) at startup and registers every priced model with its current input/output/cache rates (USD per million tokens, tiered rates where applicable).
• Pricing is never hardcoded — rates come from the catalog and are cached with a 15-minute TTL.
• Cost reporting uses catalog rates, which are the same rates the provider bills at.

Graceful fallbacks

The plugin never breaks a run or pi startup:

| Situation | Behavior | |---|---| | CHEAPERINFERENCE_API_KEY unset | Provider not registered; warning printed with signup link | | Catalog fetch fails / invalid | Provider not registered (or stale cache reused mid-session); warning printed | | Unknown model requested | No pricing invented; run proceeds |

Optional overrides: CHEAPERINFERENCE_BASE_URL (default https://api.cheaperinference.com).

Not yet

• Settled-cost precedence: CheaperInference returns the exact billed amount in response usage metadata; overriding pi's local estimate with it needs a custom streaming shim and is deferred (catalog rates are the same numbers the provider bills at).
• Reasoning/thinking metadata: models register with reasoning: false; override per-model in models.json if a catalog model supports thinking.

Tests

node --test test/

Covers catalog parsing (all documented envelope shapes), alias/normalized model matching, pricing fallback, TTL caching, and registration behavior on missing credentials and fetch failure.