TokenTicker

A price-per-million-token index for major Western & Eastern LLM API providers — Anthropic, OpenAI, Google, xAI, Meta, Mistral, DeepSeek, Qwen, Moonshot, and Z.ai.

How freshness works

There is no canonical cross-provider live pricing API, so prices are pulled from the OpenRouter models feed (/api/v1/models), which reports the actual per-token price each provider bills. Two independent mechanisms keep the page fresh:

1. Scheduled auto-update (GitHub Action, daily). .github/workflows/update-prices.yml runs update.mjs → fetches the feed → validates (sanity bounds + rejects any change larger than ±40%, keeping the previous value and flagging it in the run summary) → rewrites pricing.json → appends a dated snapshot to history.json → rebuilds the HTML → commits. Nothing implausible ships silently. Works everywhere, including the static Artifact snapshot.
2. Runtime fetch (Pages/standalone only). On load the page fetches the same feed and refreshes the current numbers, showing a live · status; on any failure it falls back to the committed pricing.json and shows snapshot ·. The Artifact build can't do this (its CSP blocks external requests) — it always shows the committed snapshot.

Guardrails are the same in both: bounds-check, ignore absurd jumps, keep last-known-good.

Trends / history

Every update appends a {date, prices} snapshot to history.json. The Trends view graphs any metric (input, output, or cost-per-call) over time, per model, with log/linear scale. To keep 30 series manageable, the Show row filters by the computed tags: a tag combobox that opens on the best models (Frontier + Flagship), a Repriced preset (plain-click plots all repriced models; Ctrl/⌘-click toggles a filter that narrows your current selection to its repriced members, tags left intact), and Default / All / Clear — over a provider-grouped legend with per-provider select-all and a live "N of 30 shown" count. End-labels show full model names and declutter into a clean single column; the plot grows taller as you add series, so nothing overlaps or truncates. The chart is a dependency-free inline SVG — no chart library — so it works offline, on Pages, and inside the Artifact. History starts with a single real snapshot and fills in one point per day; past points are never fabricated.

Tags

Every model carries a row of computed tags — derived from its price/context numbers, not hand-assigned, so a tag can't drift from the figure it describes (change a price and its band, ratio, and cache tag recompute; the live browser refresh re-tags too). The lone exception is Tier, the model's position in its provider's own lineup. Hover any pill for its exact rule. The facets and cut-offs:

Facet	Values (rule)
Tier	Frontier · Flagship · Workhorse · Small — curated lineup position
Price band	Premium ≥ $15 · Standard $4–15 · Value $1–4 · Budget < $1 — blended input+output, $/1M
Context	1M ctx ≥ 1M · Long 200K–1M · Standard < 200K — context window (tokens)
Cache	Deep ≥ 10× · Some < 10× · None — cached-input discount vs fresh input
Output ratio	Steep ≥ 5× · Moderate 3–5× · Flat < 3× — output price ÷ input price

In Trends, the Show row filters by these tags — [ Tags ▾ · Repriced ] │ [ Default · All · Clear ]. The combobox is grouped by facet and opens on the best models (Frontier + Flagship); each tag is a tri-state toggle that adds or removes its models, so selections compose (every Budget model, every Deep-cache model) and combine with the per-provider toggles below. Repriced (plain-click) plots every model whose input or output actually moved across the recorded snapshots (one real signal today, growing as history accrues). Ctrl/⌘-click it instead toggles a filter layered over your current selection: the base selection and its tags stay put — the chip and the combobox light the same amber, and the filtered-out models show struck-through in the legend — while the plot narrows to just the repriced members. It composes, so you can widen the tag selection with the filter still on. Default restores the opening plot; All / Clear select everything or nothing (and clear the filter).

Files

File	Role
pricing.json	Current prices — machine-readable index + curation (names, ids, tiers, notes). Rewritten by update.mjs.
history.json	Append-only time series of dated snapshots (drives the trends chart).
update.mjs	Fetch feed → validate → write pricing.json + append history.json. Run by the Action.
template.html	The UI (styles + markup + logic). The only file you hand-edit for look/behavior.
build.mjs	Renders pricing.json + history.json into HTML. Zero dependencies.
index.html	Generated standalone page (what Pages serves).
llm-token-price-index.html	Generated body-only fragment for publishing as a Claude Artifact.
.github/workflows/update-prices.yml	Daily cron + manual dispatch.

index.html and the fragment are build outputs — don't hand-edit; they're overwritten.

Update / run locally

node update.mjs   # refresh prices from the feed + append today's history snapshot
node build.mjs    # regenerate index.html + the Artifact fragment

Curation is preserved across refreshes: update.mjs only touches numeric fields and verified dates — it never overwrites model names, tiers, or notes. A model the feed doesn't match (e.g. gemini-3.1-pro) keeps its curated value and is listed under "missing from feed" in the run report. Provider line/dot colors live in pricing.json (accent / accentDark) and are injected at runtime, so a new provider needs no template edit. The palette is validated with the dataviz skill's validate_palette.js: in file order it clears the adjacent-pair gate on both light and dark surfaces, with the worst colorblind ΔE in the 6–8 "floor" band — which is why identity is also carried by the always-visible provider name and the chart's model end-label, never by color alone. With 10 providers no palette can make every pair colorblind-distinct at once (a hard limit around 8), so when many series overlap the labels do the disambiguating.

Publish

index.html is fully self-contained (no external requests required), so it works as: a plain file you send, GitHub Pages (current), Netlify/Vercel/Cloudflare Pages, or a Claude Artifact (publish the llm-token-price-index.html fragment).

Honest caveats

• The OpenRouter feed is a maintained aggregator, not each provider's canonical API — spot-check against the source links in the footer for anything mission-critical.
• Numbers reflect actual current billable price, so temporary promos show through (e.g. Claude Sonnet 5's intro $2/$10); the per-model note explains when they end.
• The trends chart is sparse until history accrues — one point today, one more per day.