RoninForge 发布 AI Price Index：带日期与一手来源的 AI 模型 API 价格记录

RoninForge AI Price Index is a dated, first-party-sourced record of AI model API prices over time, including what each model cost on a given date. The data is CC BY 4.0, every price links the official source that proves it, and no result depends on an API key.

Part of RoninForge.org, an independent open-source workshop that keeps dated, reproducible records of the AI developer tooling economy.

Most price lists tell you what a model costs today. This one tells you what it cost on a given date, when that price was last confirmed, and links the first-party source that proves it. That provenance is the point: every value carries a verifiable source and a last_validated_at date, including historical ones.

Why this index exists, what it is for, and how it is kept honest: https://roninforge.org/data/ai-price-index/

• Data license: CC BY 4.0 (see DATA-LICENSE.md). Use it anywhere, just attribute.
• Tooling license: MIT (see LICENSE).
• Maintained by RoninForge. Source of truth, methodology, and corrections are all in the open.

Status: early. The schema, validation, and methodology are in place; coverage is being seeded flagship-first (see Roadmap). Prices can change and history is partly reconstructed; trust the confidence label and the source_url on each record, not this sentence.

What is in here

schema/        JSON Schemas: the price record, the published per-model series, the index
examples/      illustrative-but-accurate records + series that the validator checks
data/          the published dataset (generated from the system of record; appears as coverage lands)
tools/         validate.mjs, a zero-dependency validator (run in CI and locally)
METHODOLOGY.md how prices are sourced, confirmed, dated, and corrected
CONTRIBUTING.md how to report or fix a price (a first-party dated source is required)

The price record

Every price is one record. The human-facing shape (what you contribute and what the validator checks) is defined in schema/price-record.schema.json. In short:

announced_change never changes what you read. price_usd on an open record is always what the model costs today, so an announced rise rides alongside it rather than replacing it. It carries the vendor's own wording verbatim, and validation fails once its date arrives, so an announcement cannot sit there unreconciled after it takes effect.

The published, served form is one file per model (schema/model-series.schema.json) plus a thin index.json (schema/index.schema.json); a client binary-searches a model's interval array by date.

Use the data

Three ways, depending on what you are building.

npm package (point-in-time lookups, no network)

npm install ai-price-index

The dataset ships bundled inside the package (no runtime network call), so a lookup is deterministic and reproducible against the version you installed.

import { current, priceOn, rate, meta } from 'ai-price-index';

current('claude-opus-4-8');        // today's input/output rate + the source that proves it
priceOn('gpt-4', '2024-01-01');    // the rate that was IN EFFECT on that date ($30 / $60 per Mtok)
rate('gpt-4', '2024-01-01');       // -> { provider, model, date, inputPerM: 30, outputPerM: 60 }
meta.dataModified;                 // the dataset date this install pins to

Short ids and a trailing [1m] display suffix resolve to the dated series (e.g. claude-opus-4-5 -> claude-opus-4-5-20251101); pass { provider } to disambiguate a bare model id. usdForRollup values a token rollup at a point in time with the shared cache multipliers (read 0.1x, write 1.25x / 2x). The package reproduces the same golden vectors as the other RoninForge pricing engines.

CLI (no install needed):

npx ai-price-index claude-opus-4-8                 # today's rate
npx ai-price-index gpt-4 --on 2024-01-01           # a past date
npx ai-price-index list --provider openai          # known models
npx ai-price-index claude-opus-4-8 --json          # machine-readable

JSON API (fetch the current snapshot)

The current prices are served as JSON with permissive CORS, so you can fetch them from a browser or a script. These are stable URLs (safe to depend on):

https://roninforge.org/data/ai-price-index/current.json   # current price of every model
https://roninforge.org/data/ai-price-index/index.json     # model list + per-model series file map

const { prices } = await (await fetch('https://roninforge.org/data/ai-price-index/current.json')).json();
const opus = prices.find((p) => p.model === 'claude-opus-4-8' && p.variation === 'input');
console.log(opus.price_usd, opus.src);  // 5  https://www.anthropic.com/news/claude-opus-4-8

For full price history (every dated interval, not just current), use the npm package or clone the repo. Per-model series filenames under models/ are content-hashed and not stable; do not hardcode them.

Clone the raw dataset

git clone https://github.com/RoninForge/ai-price-index
node tools/validate.mjs        # validate everything locally

When you publish anything derived from this data, attribute it: "AI Price Index by RoninForge, CC BY 4.0" with a link to https://roninforge.org/data/ai-price-index/ and the validation date of the values you used. See ATTRIBUTION.md for copy-paste formats.

How to cite

Prices change and corrections supersede without deleting history, so cite a dated release to keep your snapshot reproducible. Machine-readable citation metadata is in CITATION.cff (GitHub renders a "Cite this repository" button from it).

Plain:

AI Price Index by RoninForge (https://roninforge.org/data/ai-price-index/), CC BY 4.0. Release <tag>, accessed <date>.

BibTeX:

@misc{roninforge_ai_price_index,
  author       = {{RoninForge}},
  title        = {{AI Price Index: dated, first-party AI model API prices over time}},
  howpublished = {\url{https://roninforge.org/data/ai-price-index/}},
  note         = {Release <tag>. Data licensed CC BY 4.0. Accessed <date>.}
}

For academic reuse, cite the DOI (Zenodo): 10.5281/zenodo.20730240. This is the concept DOI and always resolves to the latest version; each release also gets its own version DOI for an exact, reproducible snapshot.

Verify a release (signed interval history)

current.json is a flat snapshot of today's prices. It carries no effective dates, so it cannot price historical usage: a consumer would have to either reprice all history at today's rate or invent a boundary date. For that, each release also publishes the full half-open [from, to) interval history for Anthropic as a single file, with a detached signature:

https://roninforge.org/data/ai-price-index/history/anthropic.json
https://roninforge.org/data/ai-price-index/history/anthropic.json.minisig

The signature is minisign format (Ed25519). Do not take our word for the numbers, check them:

curl -fsSLO https://roninforge.org/data/ai-price-index/history/anthropic.json
curl -fsSLO https://roninforge.org/data/ai-price-index/history/anthropic.json.minisig

minisign -V -m anthropic.json \
  -P 'RWRFfFveQwl6NGYtfQNtpgdecGGC3U8k5iqK+vGmUq3D3SP0wIfmg8P1'

A successful check also prints the trusted comment, which names the dataset release and its data date and is itself covered by the signature:

Signature and comment signature verified
Trusted comment: ai-price-index v2026.07.30-abc1234 dataModified=2026-07-29

Public key RWRFfFveQwl6NGYtfQNtpgdecGGC3U8k5iqK+vGmUq3D3SP0wIfmg8P1 (key id 347A0943DE5B7C45).

Signing happens in this repository's CI, and the private key exists only as a secret here. It is deliberately not on the web server. That server terminates TLS with a valid certificate, so TLS proves bytes came from that host, not that they are genuine; a host that both served and signed the data would produce perfectly valid signatures after a single-box compromise. Keeping the key out of the serving path means the web server is an untrusted mirror, and a tampered file simply fails to verify.

What the signature does not cover, stated plainly: a price that was recorded wrongly in the first place. It proves the file is the one we published, not that we published the right number. Prices carry a confidence field and a first-party src for exactly that reason.

Roadmap

Coverage is seeded flagship-first, deep on history: the leading models (Claude, GPT, Gemini) with their price history pushed back as far as it can be verifiably sourced, before widening to the long tail. Reconstructed history is labelled by confidence (archived / inferred / estimated) and never rendered with the same weight as a live-confirmed price.

Contributing

Found a wrong or missing price? Open an issue or a PR with the correct value and a first-party source URL plus the date you saw it. A contribution without a dated first-party source cannot be accepted; that rule is what makes this trustworthy. See CONTRIBUTING.md.

Docs

• The dataset, the current price table and why it exists
• Query it from Claude Code or Cursor over MCP
• Why a price change should not restate your recorded history
• AI prices did not fall, they fanned out
• Every dated price change, newest first
• Per-model pricing pages with dated history

The AI Price Index is part of RoninForge.org.