Prism

Know what an LLM call costs before you make it.

Prism is Hermedion's open-source LLM cost engine. Point it at a prompt and it tokenizes the input, prices it against a cross-provider model catalog, and ranks every model by estimated cost — before a single token is sent. It is the first building block of a gateway focused on the two things existing routers treat as afterthoughts: upfront price estimation and observability.

Why

Routers like OpenRouter tell you what a request cost after the money is spent. Prism inverts that:

• Estimate first. Per-model input/output cost for your exact prompt, cheapest first, with context-window fit checks.
• Compare instantly. One command shows the spread between the cheapest and priciest model for the same request — often 99%+.
• Build on it. The estimator is a typed, dependency-light library you can embed in CI budgets, request guards, or routing logic.

Quickstart

pnpm install
pnpm build

# List the model catalog with pricing
node packages/cli/dist/index.js models

# Estimate a prompt across every model, cheapest first
node packages/cli/dist/index.js estimate "Summarize this contract" -o 800

# Restrict to specific models, emit JSON, or read from a file/stdin
node packages/cli/dist/index.js estimate -f prompt.txt -m openai/gpt-4o,anthropic/claude-sonnet-4 --json

Example output:

MODEL                       IN TOK  OUT TOK  EST COST   VS PRICIEST
openai/gpt-4o-mini          16      800      $0.000482  -99.2%
deepseek/deepseek-chat      18      800      $0.000885  -98.5%
...
anthropic/claude-opus-4     19      800      $0.0603    —

Cheapest: openai/gpt-4o-mini at $0.000482

Library usage

import { estimate, cheapest } from "@hermedion/prism-core";

const results = estimate({
  input: [{ role: "user", content: "Summarize this contract" }],
  expectedOutputTokens: 800,
});

const winner = cheapest({ input: "Summarize this contract" });
console.log(winner.model.id, winner.totalCostUsd);

How estimation works

• OpenAI-family models are tokenized exactly with their published BPE encodings (o200k_base / cl100k_base).
• Providers without public JS tokenizers are counted with the closest encoding and corrected by a per-model factor, keeping estimates within a few percent.
• Chat inputs include per-message formatting overhead, matching how providers bill in practice.
• Pricing is a versioned snapshot (PRICING_AS_OF); verify against provider pricing pages before making billing decisions.

Packages

Package	Description
@hermedion/prism-core	Model catalog, tokenizer, and cost estimator
@hermedion/prism-cli	prism command-line interface

Roadmap

• Live pricing sync from provider pricing pages
• OpenAI-compatible routing proxy with cost-aware fallbacks
• Per-key spend analytics, traces, and latency dashboards
• Budget guards: reject or reroute requests that exceed a cost ceiling

Development

pnpm install
pnpm build
pnpm test

License

MIT © Hermedion