Skip to main content

Billing & Credits

Fund AI usage with pay-as-you-go credits. Primex also implements windowed subscription plans, but new plan sales are currently paused. Existing active plans continue to use their allowances. Both billing modes use the same prx- key and usage ledger; checkout shows current purchase availability.

How metering works​

  • Post-paid, per request. Credit moves only after successful work and is sized from reported tokens or measured service units.
  • Outcome-based units. Services may use tokens, images, megapixels, video seconds, audio minutes, characters, pages, text records, searches, or enabled tool calls. Historical ledger rows can retain older unit types.
  • Failed requests aren't billed. Successfully delivered usage is metered. Backend-unreachable, upstream-rejected, and invalid-response requests with no delivered usage settle at $0.

For token-priced models, the request ledger keeps total input, output, cache-read, and cache-write tokens as separate buckets. Cache buckets are subsets of total input, not extra tokens to add to it. Primex applies local token arithmetic only when the catalogue has an explicit rate for every bucket the provider reported, including the applicable long-context tier. If a successful provider response omits public usage or reports a bucket whose rate is unavailable, Primex uses the provider adapter's positive whole-request cost when it is safe to do so. If neither meter is available, the request settles at $0 and that model is temporarily gated while its meter recovers. Request or response byte size is never used as a final charge.

The same completeness rule applies to audio-token and image-token sub-buckets, and native-unit models must have their unit price configured before provider work starts. Their successful responses must also expose the actual page, image, video, record, character, transaction, or call count. Primex does not replace a missing meter with a one-unit default, requested duration/count, or inferred token split. Megapixel-priced images require explicit numeric dimensions, and image moderation requires its image-specific rate rather than inheriting the text rate. A missing rate or unit meter is never silently treated as zero or estimated into a final charge.

Long-running Primex Create image and video jobs are the reservation exception. Primex can temporarily reserve estimated wallet capacity while a job runs to prevent concurrent jobs from overspending. A reservation is not a debit. It is released on every exit, failed jobs settle at $0, and successful jobs settle from actual measured usage. Windowed plans do not use wallet reservations.

Broken-stream protection​

A stream can open with 200 and then carry only a provider error, without usable output. Primex detects that case and settles it at $0 rather than billing an estimate. If usable output was successfully delivered before a later stream error, the delivered usage may be metered. A genuine reply that merely contains the word "error" is unaffected.

Refunds​

If a charge still slips through incorrectly, Primex support can issue a refund to your paid balance. Refunds are idempotent (you can't be double-refunded) and are logged. Usage shows the request ID, source, key, served model, measured units, duration, status, and charge. Use the request ID when contacting support about a charge.

Add AI credits​

Open Billing in the Browser App and choose Add funds, or buy credits under My AI → Buy credits in the Telegram Mini App. Telegram is not required for browser purchases. Both fund the same AI balance when you use the same linked account.

Enter a custom USD amount or choose an amount shortcut. Checkout shows the current minimum, maximum, any eligible top-up uplift, and the exact credit you will receive. An uplift on purchased credit is paid AI credit; promotional bonus credit is a separate balance. Credit becomes available after payment settlement is confirmed.

Subscriptions (flat-fee plans)​

When sales are enabled, Primex offers flat-fee plans with rolling usage allowances instead of a pay-as-you-go wallet. Each plan grants a 5-hour and a 7-day rolling usage window, plus a concurrency limit (how many requests run at once). Plans scale from light to heavy usage. Allowance becomes available as earlier requests age out of each window; it is not a fixed hourly or daily reset.

When available, plans, prices, allowances, and concurrency appear under My AI → Plans in the Telegram Mini App. Browser Billing shows an existing account's billing mode and supports credit top-ups; it has no plan-purchase flow. Client configuration does not change between billing modes.

Which one for coding agents?​

Coding clients use the same prx- key with credits or an existing active plan. Agent sessions can consume substantial tokens, so monitor Usage and consider a monthly key spend cap. See Client Setup.

Pricing​

This page doesn't list numbers. Per-model billing units are on the Models page. The live Catalogue shows current customer prices. Billing and checkout show your balance, funding rules, purchase availability, and the final purchase quote.

Payment​

Checkout supports crypto and, when configured, hosted card, bank, digital-wallet, and local payment methods. The methods available for your order are shown at checkout.

Reading the usage ledger​

The 1D, 7D, and 1M selector controls totals, chart data, and the service breakdown. Recent Requests is the complete paginated history. The token chart only describes token-based activity; use the spend chart to compare all capability types. Router rows name the model actually served, and built-in tools appear as separate service usage where applicable. Token request rows show nonzero cache-read and cache-write buckets, and request details identify whether the final meter came from provider-reported usage, provider-adapter cost, or measured service units. Historical charge totals are read from the ledger rather than recalculated from today's catalogue prices. Rows created by the retired payload-size estimator are explicitly labeled as legacy estimates.