statlyte — live LLM pricing

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: 'per_minute' | 'per_million_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: list_models, get_model_pricing, estimate_cost, cheapest_for_workload, scheduled_price_changes.

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

Or just take the JSON

No install, no key, CORS open:

https://statlyte.com/api/v1/models
https://statlyte.com/api/v1/models/anthropic/claude-opus-5
https://statlyte.com/api/v1/changes

The raw dataset also lives in this repo as models.json and changes.json, updated by commit — so you can diff it, pin it, or vendor it.

Where the numbers come from

A job re-reads each provider's published pricing page every three hours. When a figure differs from the last one on file it writes a new observation with a timestamp and the URL it was read from. Nothing is inferred and nothing is estimated: if a price isn't published, it isn't listed.

Two honest caveats:

1. These are list prices. Negotiated, enterprise, regional and committed-spend rates differ, sometimes a lot. Confirm with the vendor before making a commercial decision.
2. Cheaper is not the same as substitutable. This records what models cost, not what they can do. rankByCost will happily tell you an 8B model is cheaper than a frontier one. That is arithmetic, not advice.

Found a figure that disagrees with a vendor's page? The vendor is right and we're wrong — open an issue and it gets fixed on the next run.

The rest of it

statlyte.com has the human-facing side: a change log, a calculator that puts two models head to head at your own volume, and a calendar of announced changes. Free, no account.

MIT licensed. Attribution appreciated, not required.