AlexRyan92/llm-price-data：把 LLM API 价格做成可验证的代码数据

Hand-verified LLM API prices as code. Dated, schema-validated, with a structured change log: the data provider pricing pages refuse to be.

Provider pricing pages have no changelog, no history, and no feed. This repo gives the table all three:

• data/models.json: USD per million tokens, standard tier, with context windows and long-context surcharge cliffs. Every number read directly from the provider's public page, with the verification date attached.
• data/events.json: a structured event per price change, model launch, or verification pass. No event, no noise. Zero price changes since tracking began (Sep 3, 2026) is itself recorded data.
• CI schema validation, so a malformed PR cannot land.

The same data drives the hosted price tracker (sortable, exportable) and its RSS feed, which fires only when a price actually moves.

Use the data directly

curl -s https://raw.githubusercontent.com/AlexRyan92/llm-price-data/main/data/models.json | jq '.models[] | {id, in, out}'

CLI

$ npx llm-price-data            # or: llm-prices after npm install -g
verified Oct 7, 2026 | blended at 10:1 input:output

model                        provider     $in/M    $out/M  blended/M
GPT-6 Luna                   OpenAI       $0.1      $0.5      $0.14
Gemini 3.1 Flash-Lite        Google       $0.25     $1.5      $0.36
Claude Haiku 4.5             Anthropic    $1        $5        $1.36
...

$ npx llm-prices --events
2026-10-07  [verify]  Initial publication. All 11 model rates verified...

--ratio 3 for chat-shaped workloads, --csv for spreadsheets, --validate for CI.

Library

import { loadModels, blended, rank, validate } from 'llm-price-data';

const { models, verified } = loadModels();
blended(models[0], 10);          // $/M for a 10:1 input:output workload
rank(models, 20, { providers: ['Anthropic'] });
validate();                      // [] when the data files are well-formed

Blended cost is (ratio * in + out) / (ratio + 1). Long-context surcharges and cache discounts are deliberately not blended in; they live on the model entries (longCtx) for your own math.

Corrections and updates

If a number here disagrees with a provider's page, the provider's page wins and I owe the fix. PRs welcome with one rule: link the provider pricing page you read the number from. Every merged change lands as an event in events.json, which is what makes the history trustworthy.

License

Code MIT, data CC BY 4.0 © Alexander Ryan.