Productizing

Pricing Models for AI Features: Usage, Credits, BYO-Key, or Managed

⏱ 15 min

Every AI feature has a real, variable cost behind it: a call to Claude, GPT, or Gemini through the PHP AI Client SDK costs the provider money per token, and someone has to pay that bill. Traditional WordPress plugins mostly don’t have this problem, a form builder or an SEO plugin costs you the same to run whether one customer uses it or ten thousand. An AI feature doesn’t work that way, and the pricing model you choose is really a decision about who carries that variable cost and how much friction you’re willing to add to shift it. This lesson lays out the four common patterns honestly, without pretending one is universally correct.

What you'll learn in this lesson
The four common models
Usage-based, credits, bring-your-own-key, and managed/flat-fee, and what each actually commits you to.
Who bears the AI cost risk in each model
The single most important variable in this decision.
Setup friction versus cost predictability
Why the easiest model to sell isn't always the easiest to operate.
Hybrid approaches
Why most real products end up mixing these rather than picking exactly one.
Prerequisites

This lesson assumes you’re pricing a real feature, not a hypothetical, and that you’ve already worked through Lesson 1’s productizing checklist. Nothing here requires new code, this is a business decision that then shapes the licensing system in Lesson 3.

Step 1: The four models, compared honestly

Pricing models for an AI-powered WordPress feature
ModelHow it worksWho bears AI cost riskSetup friction
Usage-basedCustomer pays you per action or per unit of usage (per call, per generated item), you pay the AI provider and take a marginYou, until you set pricing high enough to absorb varianceLow, customer just pays and uses it
CreditsCustomer buys a prepaid bundle of credits that abilities consume at defined rates, you pay the provider as credits are spentYou, but bounded, since credits are prepaid and revenue arrives before the cost doesLow to moderate, customer needs to understand what a credit “costs” in real usage
Bring-your-own-key (BYO-key)Customer supplies their own API key for Anthropic, OpenAI, or Google, your plugin calls the provider directly with the customer’s key and credentialsThe customer, entirely, they pay the provider directlyHigher, the customer has to create a provider account, generate a key, and manage their own billing with that provider
Managed / flat-feeCustomer pays one flat license fee (one-time or subscription) regardless of how much AI usage they generate, you cover all provider costs from that revenueYou, fully, with no per-use ceiling unless you build usage limits into the product itselfLowest, the customer pays once and it just works

This is the same “who bears the cost, and how much friction does shifting it create” tradeoff that shows up throughout SaaS and API-driven products generally, applied here to a WordPress plugin specifically.

Bring-your-own-key shifts cost risk entirely onto the customer: they create their own Anthropic, OpenAI, or Google account, generate a key, and every API call your plugin makes on their behalf bills directly to their own provider account. This is exactly the pattern the PHP AI Client SDK is built around (Course 5 covered the setup in detail), and it’s a natural fit for a plugin sold to developers or technical site owners who are already comfortable managing an API key. The tradeoff is setup friction: a customer who has never used an AI provider’s console before now has an extra step between “installed the plugin” and “saw it work,” which is exactly the kind of friction Lesson 1 flagged as a real onboarding risk. It also means your own margin isn’t tied to AI cost at all, you’re purely selling the plugin’s functionality, workflow, and interface, not reselling tokens.

A managed model removes that friction entirely: no key to create, no provider account, one price, it works. This is the model most non-technical WordPress users expect, since it matches how they already buy every other plugin or theme. The real exposure is yours: if a customer runs an unusually high volume of AI calls (or if a bug in your plugin causes runaway calls, exactly the kind of failure Course 4’s rate-limiting patterns exist to prevent), you’re paying the provider bill for that regardless of what the customer paid you. A managed model without usage limits, or without the underlying rate limiting already built for security reasons, is a real financial risk, not just a security one.

A managed model needs the same rate limiting a secure plugin needs anyway

If you followed Course 4’s custom rate-limiting patterns because the MCP Adapter doesn’t ship them, you already have most of what a managed pricing model needs to stay financially safe. Rate limiting isn’t just a security control here, it’s a cost control, and a managed-pricing product without it is exposed to unbounded usage from a single customer or a runaway agent loop.

Step 4: Usage-based and credits sit in between

Usage-based pricing (pay per call, or per generated item) and credit bundles (prepay for a set amount of usage) both let you pass AI cost through to the customer while keeping the AI provider relationship on your side, so the customer never touches an API key. Credits in particular give you a natural, visible unit customers can reason about (“this bundle covers roughly this many summaries”), and they front-load revenue before the cost is incurred, which is friendlier to your cash flow than pure usage-based billing where you cover the AI cost immediately and invoice or charge later.

Questions that actually decide this for your product
Is your audience developers/agencies, or non-technical site owners?
BYO-key fits the former far better than the latter.
How predictable is your per-action AI cost?
Wildly variable cost (long-context RAG queries versus short completions) favors usage-based or credits over flat-fee.
Do you already have rate limiting in place?
If not, a managed model is riskier than it looks until you build it.
Can you support multiple models at once?
Many real products offer a managed tier and a BYO-key tier side by side, rather than forcing one choice on every customer.

Recap

There’s no universally correct pricing model for an AI feature, only a real tradeoff between who bears the variable AI cost and how much setup friction the customer accepts. BYO-key shifts cost to the customer at the price of setup friction and fits technical audiences well. Managed/flat-fee removes friction entirely but exposes you to variable cost and needs the rate limiting Course 4 already gave you good reasons to build. Usage-based and credit models sit between these two extremes and are common patterns for products selling to a broad audience. Many real products end up offering more than one tier rather than forcing every customer through the same model.

Resources & further reading

← From Working Plugin to Sellable Product Licensing and License Key Systems →