promptspend

https://github.com/user-attachments/assets/6bfecd10-84a9-408b-b08c-ff9481648692

Why this exists

Every LLM cost calculator on the web has the same failure mode: it is a snapshot. Someone builds it, hard-codes a dozen model prices, and within a few months the entire premise is wrong — the models it compares have been superseded and the prices it quotes no longer exist.

PromptSpend is built the other way round. The pricing pipeline is the product; the calculator is what sits on top of it. Every morning a GitHub Action re-fetches the catalog from independent sources, merges them under an explicit trust order, runs sanity checks, and either commits the result or opens a pull request for a human. Capture patterns are family-level, so a brand-new model version is picked up automatically without anyone touching code.

Related MCP server: Model Price Watch

What it does

• Receipt — copy a fully visible, one-response cost-audit object into an existing AI conversation. It estimates the visible workload, retrieves current PromptSpend prices, names the two largest cost drivers and suggests cheaper models worth testing without pretending price proves equal quality.

• Estimate — describe one interaction (paste your real prompt or move the sliders), set your scale, and see the monthly, yearly, per-user and margin numbers for up to four models side by side.

• Compare — every tracked model on a price-versus-capability value map, plus a sortable catalog you can select from directly, with the source and verification date for every row.

• Learn — seven short interactive lessons on tokens, why output costs more, how wide the price spread is, compounding chat history, hidden reasoning tokens, caching and batching, and how the data pipeline works.

• Free tiers — which providers let a new account start without paying, card rules, limits and whether free use trains their models, quoted word for word from each vendor's own pages with a link and date on every fact, and re-checked each morning. One comparison page and one page per provider; see docs/FREE_TIERS.md.

• Data & Alerts — pipeline health, full provenance for every number with a link to the vendor page it came from, what is currently flagged, and four ways to hear about a price change: an Atom feed, the repository's own pull requests, browser push, and an email digest.

Price alerts

Opt-in, and off unless a deployment is configured for them. Browser push stores nothing personal — a push subscription is an opaque URL the browser issues. Email is double opt-in with one-click unsubscribe, and stores your address, the models you follow, and the date you asked; nothing else. Either channel can watch the whole catalog or just the models you pick.

The delivery service is a Cloudflare Worker in worker/, with the push payload encryption written out against RFC 8291 and checked byte for byte against the RFC's own worked example. docs/ALERTS.md has the architecture, the cost model and the domain cutover.

Things it does that most calculators get wrong

How the data stays current

                    ┌─────────────────────────┐
  LiteLLM catalog ──┤                         │
                    │   allowlist → merge →   │──→ public/data/pricing.json ──→ the app
  OpenRouter API ───┤   validate → diff       │
                    └───────────┬─────────────┘
  data/pricing-overrides.json ──┘            └──→ docs/pricing-changelog.md
       (hand-verified, wins)

The trust ladder, in order:

1. Hand-verified vendor rates (data/pricing-overrides.json) — win every conflict.

2. The LiteLLM community catalog — the automated daily feed; thousands of models, updated within days of a release.

3. The OpenRouter API — an independent cross-check, never a source of record. Reseller pricing differs from first-party list pricing, so a disagreement of more than 20% flags the model rather than changing it.

4. Sanity rules — schema validation, non-negative rates, implausibility checks, a review flag on any price that moved more than 50% in a single day, a floor on how many rows a source may return, and a cap on how much the catalog may shrink in one run.

A diff is committed and deployed automatically, flags included: there is no human gate on publication, on purpose, because a flag that held the whole catalog once left the site's dates two days behind. A newly raised flag publishes with everything else, marks that row disputed on the site, and opens an issue for review — a long-standing disagreement does not re-open one every morning. A run that loses a source or trips a size guard is degraded: it publishes nothing, records why in public/data/sync-status.json, and fails loudly. Either way the change lands in docs/pricing-changelog.md.

Nothing disappears on one bad morning. A model missing from the feed is kept and marked stale, not deleted; retiring one for good is a deliberate edit to data/models-allowlist.json.

Rung 1 is worked by hand, and the split is published rather than blurred. Every row says which rung it came from, and rows read against a vendor's own page carry the URL that was read and the date it was read. Most of the catalog is now vendor-sourced — but not all of it, because some vendors do not publish a pricing page that lists their own models, and the honest answer there is to leave the row labelled litellm rather than launder an aggregator's number into a first-party claim. The models this currently affects are named in docs/DEFERRED.md. A number is only worth what its worst source is, so the site shows you the source per row instead of an average you cannot inspect.

Something watches the promise

A daily job reads the live site — not the copy in this repository — and raises an issue if the published catalog is more than two days old. That distinction is the whole point. Between a price moving upstream and you seeing it, four things can fail: the sync errors, the sync cannot commit its result, the deploy fails after a green sync, or the CDN serves a stale artifact. Checking the file in git catches the first of those. Fetching what the site actually serves catches all four.

It exists because the sync once failed at its publishing step and nothing said so; the site quietly served a day-old catalog until someone happened to look. A project that claims its numbers are never a year out of date should be the first to know when they are.

Two dates, not one

The site shows prices last changed and sources last checked separately, because they answer different questions. A quiet week in the market makes the first date age while everything works perfectly; a broken scheduler looks exactly the same if you only publish one date. The second comes from a health manifest written on every run, successful or not:

https://promptspend.com/data/sync-status.json

Right now the first of those reads not yet recorded, and that is correct rather than broken. No vendor has moved a price since this catalog's history began on 2026-08-01, so there is no date to show — and the alternative, quietly displaying the last build date, would announce a price change every morning nothing happened. An empty field you can trust beats a populated one you cannot.

The second is now a state and not just a date: the chip on the results panel is green while the sources were checked within a day, blue at two days, amber past the window the freshness monitor tolerates or after a degraded run, and outlined-grey when the manifest could not be loaded at all. That last state exists on purpose. A status light that reads "fine" when its own evidence is missing is precisely the failure the monitor was built to catch.

Three things a number can do

Keeping "prices last changed" honest means the pipeline has to tell apart three events that all look like a diff:

Only the first is news. Conflating them is not hypothetical: reading x.ai's real pricing page instead of its model list once gained a long-context tier and restated a cached-input rate, and the catalog stamped both as price changes — putting PRICES CHANGED 2026-08-06 on the busiest panel of the site with no vendor having touched a rate. Suppressing corrections can hide a genuine same-day move, and that trade is deliberate: on a catalog whose whole claim is provenance, announcing a change nobody made costs more than missing one by a day.

Use the data yourself

There is a free, keyless, CORS-open API at promptspend.dev — no account, no key, no rate limit:

curl https://promptspend.dev/v1/prices          # flat rows: the numbers only
curl https://promptspend.dev/v1/models/gpt-5    # one model, in full
curl https://promptspend.dev/v1/prices.csv      # the same rows, for a spreadsheet

Filters: ?provider=, ?status=, ?aliases=include. Every response carries X-PromptSpend-Generated-At, and OpenAPI 3.1 lives at /openapi.json. See docs/API.md.

Or drop a price badge in your own README

[![claude-opus-5 price](https://promptspend.dev/badge/claude-opus-5.svg)](https://promptspend.dev/)

Swap the model id for any id from /v1/models. The badge is drawn fresh on every fetch — no cache — so the date it shows is always the day that price was last confirmed.

Or inside your coding agent

claude mcp add promptspend -- npx -y @promptspend/mcp

Other pricing MCP servers exist and index more models than this one. What none of them do is tell you where a number came from and when it was last confirmed — the competing server's own documentation says only that pricing "is updated regularly", with no verification date per price. That gap is the reason this exists.

It matters more to a model than to a person. Someone reading a web page sees the interface around a figure and calibrates. A model handed a bare number repeats it with whatever confidence the sentence implies. Given the source and the date it can say "as of 1 August, per OpenAI's pricing page" — and when two sources disagree it is told so, rather than being handed a number somebody picked.

Three tools, deliberately. estimate_cost is the one a price lookup cannot provide: it runs this repository's cost engine, imported rather than reimplemented, so it accounts for compounding conversation history, cache writes, long-context tiers and reasoning tokens — and cannot report a figure that disagrees with the calculator, because it is the same code. A test asserts that.

And because MCP tool definitions load into your context on every turn, the server's own overhead is measured, budgeted at 900, checked in CI and published: ~700 tokens. Nothing else in the ecosystem appears to publish its own footprint. See mcp/README.md.

Or in your editor

The same catalog, on the line of code that chooses the model.

response = client.messages.create(
    model="claude-sonnet-4-6", # $3 / $15 per M · max out $0.061
    max_tokens=4096,
)

A linter for model choice rather than a calculator in a sidebar — a calculator in an editor is only this website with worse ergonomics. What an editor can do that a web page cannot is notice things: a nearby max_tokens becomes the output ceiling in money, a cap above what the model can emit is named rather than priced as though it would work, deprecated and disputed rows reach the Problems panel, and an Explorer view lists every model the repository calls, dearest first.

Everything it says is Information severity, never a warning. None of these are mistakes and the extension has no idea what the code needs. The cheaper-model suggestion is off by default for the same reason the value map's capability axis is labelled illustrative: it is an opinion, where the rest are facts.

code --install-extension promptspend.promptspend

Or search PromptSpend in the Extensions pane. It is on Open VSX too, which is where Cursor, Windsurf and VSCodium look.

Same rule as everywhere else — no bundled prices. If the catalog cannot be reached it says so and shows nothing, and the status bar carries the generation date at all times. See vscode/README.md.

Or on your phone

Native apps for iPhone and Android, free on the App Store and Google Play: paste a sample conversation, compare up to four models, and see the cost per conversation, day, month and year from the same catalog. Pasted text is processed on the device and never uploaded. Both are described at promptspend.com/app; the source is in apps/mobile/.

Or read the file the API reads. The catalog is plain, versioned JSON with a stable shape:

https://promptspend.com/data/pricing.json

{
  "schemaVersion": 2,
  "generatedAt": "2026-08-02T02:27:22.781Z",
  "providers": [{ "id": "openai", "name": "OpenAI", "country": "US", "pricingUrl": "https://…" }],
  "models": [
    {
      "id": "gpt-5.6-terra",
      "providerId": "openai",
      "displayName": "GPT-5.6 Terra",
      "status": "current", // current | legacy | deprecated
      "contextWindow": 1050000,
      "pricing": {
        "input": 2, // USD per 1M tokens
        "output": 12,
        "cachedInput": 0.2,
        "cacheWrite": 2.5, // writing costs *more* than sending
        "batchDiscount": 0.5,
        "longContext": { "thresholdTokens": 272000, "input": 4, "output": 18 },
      },
      "tokenizer": { "kind": "tiktoken", "encoding": "o200k_base" },
      "provenance": {
        "source": "vendor", // vendor | litellm | openrouter
        "lastVerified": "2026-08-01", // when it was checked — always present
        "lastChanged": "2026-08-01", // optional: when the number last moved
        "verifiedUrl": "https://developers.openai.com/api/docs/pricing",
      },
    },
  ],
}

Optional fields worth knowing: aliasOf marks an id that routes to another model (so it is not counted twice), provenance.stale marks a row upstream has stopped listing, and provenance.needsReview plus reviewNote carry an unresolved disagreement and both numbers involved.

provenance.lastChanged is optional too: it is present only on a model whose rate has actually moved since this catalog's recorded history began on 2026-08-01, and absent otherwise. Read its absence as "no change on record", never as "unknown freshness": lastVerified answers that, and it is always there. A field that reported the last sync instead would read "changed today" every morning, which is the one failure a provenance catalog cannot afford — so it stays empty until a rate actually moves.

Running it locally

npm ci                 # `ci`, not `install` — this repo has a lockfile and CI honours it
npm run dev            # http://localhost:5173

npm run verify             # exactly what CI runs, and what the deploy gate runs
npm run sync:pricing:dry   # see what today's sync would change, without writing

verify is typecheck, lint, format check, an encoding check, tests with coverage thresholds, the published test counts, the published page counts, a production build, catalog schema validation, the bundle budget, the Content Security Policy and the SEO checks. The deploy workflow calls the same reusable workflow CI does and publishes the artifact it produced, so a commit that fails any of them cannot reach the live site.

It does not run the other packages' suites — CI has a job each for api/, mcp/, updates/, vscode/ and worker/. See CONTRIBUTING.md.

Project layout

src/lib/engine/     the cost engine — pure functions, no React, heavily tested
src/receipt/        the transparent Receipt object, page and versioned instruction contract
receipt/            the second Vite HTML entry, published as /receipt/
src/lib/tokenize/   exact tokenizer (lazy-loaded) + calibrated ratios
src/lib/pricing/    catalog schema, validation, lookups
src/lib/alerts/     browser-side push and alerts API client
src/components/     the four views
src/lib/seo/        the generated pages: slugs, page model, HTML renderer
scripts/            the daily sync pipeline (scripts/lib is unit-tested)
data/               capture patterns and hand-verified overrides
public/data/        the published catalog the app reads
public/sw.js        service worker — push display only, no offline cache
worker/             the alerts API (Cloudflare Worker, own package and tests)
api/                the public pricing API on promptspend.dev (own package and tests)
updates/            the app's self-hosted over-the-air update server (own package and tests)
src/state/          the scenario hook, and the URL it mirrors itself into
tests/              the browser suite: Playwright at four viewports, plus axe
tools/              the promo pipeline — capture, render, stitch
mcp/                the MCP server — imports the engine above, so it cannot disagree with it
vscode/             the VS Code extension — imports it too, for the same reason

Beyond the calculator and Receipt, the build writes 351 crawlable pages — one per model, one per provider, and a curated set of head-to-heads — from the same catalog and the same cost engine. See docs/PAGES.md.

Design and accessibility

Cobalt on a cool-paper canvas, with a distinct set of money colours that never change with branding: green means savings, red means this option costs more. The input/output chart pair is validated for colour-vision deficiency in CI — src/lib/palette.test.ts simulates protanopia and deuteranopia and fails the build if the two marks stop being distinguishable.

src/lib/contrast.test.ts reads tokens.css directly and fails the build if any accent, on any theme and any canvas, drops below 4.5:1 against a surface it can appear on. Every interactive target is at least 24×24 (44×44 on touch), no text renders below 12px, and there is no horizontal page scroll from 320px up. Keyboard navigation throughout — including every point on the value map, which is a real button — focus is returned when a dialog closes, prefers-reduced-motion is respected in JavaScript as well as CSS, and there is a Ctrl/Cmd+K command palette.

Documentation

Contributing

Adding a model is usually a one-line change. See CONTRIBUTING.md, or open a model request and someone will pick it up.

Not a code change? info@promptspend.com. Security reports go to security@promptspend.com or GitHub's private vulnerability reporting — see SECURITY.md.

The three rules that are not up for negotiation, because breaking any of them turns an estimator into a guess with good typography:

1. Never invent a rate. If a provider does not publish a number, charge full price and say so.

2. A claim on screen must be true of the code. When behaviour and copy disagree, fixing the copy is a legitimate fix; leaving both is not.

3. No enabled control that does nothing. Describe a planned feature — do not simulate it.

Honest limitations

The point of this section is that it is longer than it needs to be. An estimator that hides its edges is just a confident guess.

• Scope of the prices. Standard-tier, global-endpoint list prices in USD. Not modelled: regional and data-residency premiums (OpenAI and Anthropic both charge 1.1×), fast/priority tiers, server-side tool call fees, fine-tuning, and negotiated or committed-use discounts. The site says this under every estimate, not only here.

• "Exact" means exact raw text. The tokenizer counts the string you give it. A real request also bills message framing, tool definitions and any images. Treat an exact count as a floor.

• Token counts for non-OpenAI families are estimates from calibrated characters-per-token ratios, labelled as such everywhere they appear. Those providers do not ship a browser-runnable tokenizer.

• The capability axis on the value map is illustrative, not a benchmark. Models without an estimate are not plotted at all rather than being given a default, and the chart says how many that is.

• Caching is off by default and the estimate charges cache writes where a provider publishes a rate. Where one does not, cached tokens are billed at the full input rate rather than at an invented discount. Providers that bill cache residency separately are labelled with that per-hour rate; it is excluded from the estimate until a retention duration is known, and enabling caching raises an explicit warning.

• Long-context tiers are modelled where they are published (per request, not per conversation). Where a provider has a tier we have not recorded, the estimate says so instead of quietly using the flat rate.

• The Receipt sees a transcript, not an invoice. It cannot see hidden system prompts, provider-side tools or media, cache usage, hidden reasoning, or an exact billable model unless the conversation exposes them. It reports ranges and exclusions, and refuses to invent a dollar amount when current pricing is unavailable.

• Exact counting downloads a ~3 MB tokenizer chunk, and only when you paste text for an OpenAI-family model. The initial page is about 100 KB gzipped, and CI fails if it passes 105.

Privacy, precisely

Nothing you type or paste leaves your browser. No accounts, no ads. Fonts are self-hosted. One image is not: the footer's LaunchNest listing badge loads from launchnest.io (their directory requires it), with no referrer, so LaunchNest's server sees a visitor's IP address but not which page they were on.

The website counts visits with Google Analytics, which sets cookies. Since 2026-10-24 every page of promptspend.com loads Google Analytics 4 and shows a small notice saying so, with a link to the privacy policy and a Turn off button. Google Analytics sets two cookies (_ga, _ga_08GFGLY4VD). PromptSpend's code sends it the page's site and path only — never the query string or #fragment, where the scenario and alert tokens live — plus the page title and referring page (cut down the same way); it switches off GA's automatic page view, Google signals and ad personalisation, and passes it nothing anyone types, pastes or enters in a form (src/lib/analytics/). It runs only on promptspend.com, never when the browser sends Global Privacy Control, and one line in src/lib/analytics/config.ts (ANALYTICS_REQUIRES_CONSENT) switches it to loading only after a visitor clicks Accept. Visitors can turn it off on the notice, on the privacy page, or with Google's own opt-out add-on.

The website also keeps Cloudflare Web Analytics, a cookieless beacon that reports a page's address (without its query string or fragment), referrer, country and load timings to promptspend.com's own /cdn-cgi/rum; neither service sees pasted text. The apps, the VS Code extension, the MCP server and the API carry no analytics. Pasted prompt text is tokenised in your browser, deliberately excluded from the shareable URL, held in a bounded in-memory cache, and gone when you close the tab. localStorage holds only interface choices: your theme, accent and background, whether you dismissed the welcome and apps banners, and your answer to the analytics notice.

Price alerts are the one exception, and only if you opt in. They are a feature you have to switch on, and they are the only reason this project has a server that stores anything (a Cloudflare Worker — docs/ALERTS.md); the pricing API stores nothing but Cloudflare's request logs (kept up to seven days), and the apps' update server has its logging switched off. Precisely what changes:

• The Content Security Policy opens connect-src for that one API origin and, whenever an API is configured, script-src, connect-src and frame-src for Turnstile's challenges.cloudflare.com (whether Turnstile is actually required is decided at runtime, so the build permits it). Beyond that, the remote origins are img-src for the LaunchNest badge, script-src for the Cloudflare beacon (which posts only to this site), and Google Analytics: www.googletagmanager.com for gtag.js (in script-src, and in connect-src and img-src as Google's guidance asks), and its collection hosts (www.google-analytics.com, region1.google-analytics.com, region1.analytics.google.com) in connect-src and img-src, each named rather than wildcarded. That connect-src line is what stops a compromised dependency exfiltrating a pasted prompt, so it is generated from configured values rather than hand-maintained, and npm run check:csp fails on any origin it does not list. Admitting Google's collector is a real cost — a script that got onto the page could post to an analytics account of its own — accepted on 2026-10-06 for visit counts.

• Browser push stores nothing personal. A push subscription is an opaque URL the browser issues. No address, no name, nothing that identifies you.

• Request logs. Alert links carry a private token in their address, so the alerts Worker keeps no per-request logs (invocation_logs is off) and its own messages log counts and error types, never an address. The public pricing API keeps Cloudflare's request logs, retained for up to seven days (Cloudflare's published maximum), because its addresses carry no secrets and they are how faults get diagnosed.

• Email stores your address, the models you follow, and the date you asked. That is the whole record. No name, no raw IP (consent is recorded as a salted hash), no opens, no clicks. Cloudflare hosts the database and sends the mail; no other company receives it. Double opt-in, one-click unsubscribe, and unconfirmed addresses are deleted within a week.

• The alerts form never sees anything you paste into the estimator. Those are different parts of the page and the prompt text never leaves the browser.

A deployment with no API configured — which is what this repository builds by default — leaves the alerts API and Turnstile out of its policy (the analytics and badge origins above stay) and says on screen that alerts are not switched on, rather than rendering a form that cannot work.

Security

See SECURITY.md. Wrong prices are treated as the most serious class of bug this project can have, and are explicitly in scope for a report.

Licence

MIT — see LICENSE. The self-hosted typefaces (Space Grotesk, IBM Plex Sans, JetBrains Mono) are SIL Open Font License 1.1.