statlyte — live LLM pricing MCP server

statlyte

Live pricing, context windows and identifiers for every major LLM API — so you can stop hardcoding a model table that goes stale.

Every app that touches an LLM ends up with something like this pasted into it:

const PRICES = {
  'gpt-4o': { input: 2.5, output: 10 },
  'claude-3-5-sonnet': { input: 3, output: 15 },
  // …written once, wrong within a month
};

Then a model is retired, a new one lands, an introductory rate expires, and your cost dashboard is quietly lying to you. This package fetches the current numbers instead.

• 159 models across 13 providers — Anthropic, OpenAI, Google, xAI, DeepSeek, Mistral, Together AI, Voyage AI, Groq, Cohere, Fireworks AI, Deepgram, AssemblyAI
• Read from each vendor's own published pricing page, every three hours, with the source URL recorded
• Zero dependencies. Node, Bun, Deno, Cloudflare Workers, browser
• Bundled snapshot fallback, so a flaky network never throws in your request path
• MIT. The data is free and the API needs no key

npm i statlyte

Use it

import { getModel, costOf, rankByCost, scheduledChanges } from 'statlyte';

// What does this actually cost me?
await costOf('claude-sonnet-5', { input: 12_000, output: 800 });
// => 0.032

// Look up by statlyte id or the vendor's own API id
const m = await getModel('gpt-5-mini');
m.contextWindow;        // 400000
m.prices.input;         // 0.25  (USD per million tokens)
m.prices.cache_read;    // 0.025

// Cheapest model for a real monthly workload
const ranked = await rankByCost({ inputPerMonth: 620e6, outputPerMonth: 210e6 });
ranked[0].name;         // cheapest first
ranked[0].monthlyCost;  // USD/month

// Price rises vendors have already announced
await scheduledChanges();
// [{ name: 'Claude Sonnet 5', effectiveOn: '2026-09-01',
//    from: { input: 2, output: 10 }, to: { input: 3, output: 15 },
//    reason: 'Introductory pricing ends' }]

Everything is cached in-process for six hours. Pass { offline: true } to any call to use only the bundled snapshot and never touch the network.

Audio/transcription models (Deepgram, OpenAI Whisper/TTS) aren't priced per token — they carry m.nonTokenPrice ({ unit: 'perminute' | 'permillion_characters', amount }) instead, and m.prices is {}. costOf() throws a clear error rather than silently returning 0 if you call it on one of these; check m.nonTokenPrice first, or filter on m.prices.input != null.

Fail your build when a price is about to change

The genuinely useful trick. scheduledChanges() returns increases vendors have announced but not yet applied — so you can find out at build time rather than on the invoice:

// scripts/check-model-costs.mjs
import { scheduledChanges } from 'statlyte';

const MODELS_WE_USE = ['anthropic/claude-sonnet-5', 'openai/gpt-5-mini'];
const soon = (await scheduledChanges())
  .filter((c) => MODELS_WE_USE.includes(c.id))
  .filter((c) => new Date(c.effectiveOn) - Date.now() < 60 * 86400_000);

if (soon.length) {
  console.error('Price change coming:');
  for (const c of soon) {
    console.error(`  ${c.name} on ${c.effectiveOn}: ` +
      `in $${c.from.input}→$${c.to.input}, out $${c.from.output}→$${c.to.output} per MTok`);
  }
  process.exit(1);
}

MCP server

An assistant's training data goes stale on prices within weeks, and a guessed number is worse than no number. This gives your agent the current figures:

claude mcp add statlyte -- npx -y statlyte

Other MCP clients

{
  "mcpServers": {
    "statlyte": {
      "command": "npx",
      "args": ["-y", "statlyte"]
    }
  }
}

Tools: listmodels, getmodelpricing, estimatecost, cheapestforworkload, scheduledpricechanges.

Also listed in the official MCP Registry as io.github.richardwilkinson9/statlyte.

Or just take the JSON

No install, no key, CORS open:

Shortened. The full README is on GitHub.