Watch Claude pricing for divergent A/B variants

examples/convex-price-watch

Contributors: Nikola BalicUpdated May 4, 2026

A scheduled monitor: scrape https://claude.com/pricing every 10 minutes from two parallel proxy probes, store one row per (region, tier, capturedAt) in Convex, and surface tiers where the probes disagree. No LLM, no streaming. Just @steel-dev/convex, a cron, and a reactive dashboard.

convex/
├── convex.config.ts    mounts the steel component
├── schema.ts           priceSnapshots table
├── scraper.ts          captureFromRegion / captureAll / snapshotNow
├── crons.ts            10-minute schedule
└── prices.ts           current / history / recent / recentDivergences
src/
└── App.tsx             tiers × regions grid + divergence callout

The scraper fans out across two Steel deployment regions (lax, iad) with useProxy: true. Each scrape exits through a random residential IP from Steel's pool, so two parallel calls exercise different IPs and pick up A/B pricing experiments that depend on the visitor bucket.

Run it

cd examples/convex-price-watch
npm install
cp .env.example .env       # fill in STEEL_API_KEY
npx convex dev             # creates a dev deployment on first run

In a second terminal:

npx convex env set STEEL_API_KEY "$STEEL_API_KEY"
npm run dev                # vite frontend

Get a key at app.steel.dev/settings/api-keys.

Open the Vite URL and click Snapshot now. Within ~15-30 seconds the grid populates with the latest prices. The cron also runs captureAll every 10 minutes once the deployment is live.

Claude pricing watch
            LAX             IAD
Free        $0     just now $0     just now
Pro         $20    just now $20    just now
Max         $100   just now $100   just now

Proxy-routed via Steel. Cron runs every 10 minutes.

If the probes disagree, a yellow-bordered "Divergence detected" card appears above the grid. The cell whose amount differs from the per-tier majority gets a yellow tint.

Why two parallel probes

useProxy on ScrapeParams is a boolean. Setting it to true routes the request through a random residential IP each time, but you don't get to pick the country from the scrape API. The country-pinned form ({ geolocation: { country } }) lives on steel.sessions.create, not steel.steel.scrape.

Two probes running in different deployment regions (region: "lax" and region: "iad") give you two different IPs per tick. That's enough to surface visitor-bucket A/B variance: if Anthropic serves $17 to one bucket and $20 to another, the probes have a real chance of landing in different buckets and the grid lights up.

True per-country routing is a one-line extension: open a session with useProxy: { geolocation: { country: "DE" } }, then scrape through the session id. Listed under Make it yours.

Each probe is wrapped in try/catch inside captureAll. A 503 from one Steel region doesn't blank the others.

The delay is load-bearing

claude.com/pricing ships a small nav-only stub on the first response and hydrates the actual tier prices client-side. Without delay: 5000 on the scrape call, the probe grabs the stub, the tier regex finds nothing, and the row count is zero.

captureFromRegion retries once if the markdown is empty or doesn't contain any tier name. The retry covers the residual hydration flake on residential IPs without inflating latency on good runs.

const result = await steel.steel.scrape(
  ctx,
  {
    url: TARGET_URL,
    delay: 5000,
    commandArgs: { format: ["markdown"], useProxy: true, region },
  },
  { ownerId: "monitor" },
);

extractTierPrice then walks the markdown for each of Free, Pro, Max, finds the first mention case-insensitively, and matches ($|€|£)N in the next 600 characters. Fragile to layout changes; good enough for a known target.

Make it yours

• Country-pinned probes. Swap steel.steel.scrape for steel.sessions.create({ sessionArgs: { useProxy: { geolocation: { country: "DE" } } } }) and scrape through that session. Repeat for each country you want to watch.
• Alert on divergence. Add an HTTP action that posts to Slack or Discord whenever recentDivergences returns a non-empty array. Schedule it on the same cron, after captureAll.
• Watch more sites. TARGET_URL and TIERS are constants at the top of scraper.ts. Add a second target with its own tier list and a second cron.
• Store screenshots alongside prices. Pass format: ["markdown", "screenshot"] to commandArgs and write the screenshot to ctx.storage. Useful when a layout change breaks the regex and you want to see what the page actually looked like.
• Diff Pro/Max/Team additions. Persist the full TIERS array per snapshot and compare the latest captured set against the prior one. New or removed tiers are first-class signal.

Related

• @steel-dev/convex component
• Convex crons
• Sibling recipe: convex-chat-with-page (interactive agent, streamed)

Related recipes