Bring your own key, or BYOK, means connecting an API key from a provider account you control. The provider bills supported usage to that account. An AI subscription usually combines access to an application with its own allowances, usage rules and product features.
Neither approach is automatically cheaper. The useful comparison is the total cost of finishing your workload with the features you need: model usage, application fees, tools, media, retries and any manual work left over.
This guide explains how to compare them and includes three hypothetical cost scenarios. The scenarios demonstrate arithmetic; they are not current prices or observed bills for a named service. You can download the cost worksheet and replace the assumptions with your own data.
What each payment covers
There can be several different charges in an AI workflow:
| Payment | What it may cover | What to check |
|---|---|---|
| Application subscription | Interface, storage, collaboration and included usage | Eligible features, allowances and plan restrictions |
| Provider API usage | Model requests through your provider account | Model rates, tool fees, billing settings and usage limits |
| Additional application usage | Work beyond included allowances or separately priced tools | Estimate, maximum charge and funding source |
| Specialist services | Editing, distribution, hosting or other work | Whether a separate account or subscription is needed |
A BYOK application can still charge for the application itself. A subscription can still meter expensive work. “Included access” and “unlimited use” are not interchangeable descriptions.
Read the terms of the specific route you plan to use. A direct provider key, a reseller's included allowance and a connected third-party account can have different billing and access rules even when they expose a similarly named model.
Separate consumer subscriptions from API billing
A provider's consumer app and its API are different products. Do not assume that paying for one automatically funds every request from an external application.
For example, OpenAI maintains separate ChatGPT and API billing systems, while eligible subscriptions may have specific API benefits. Check your actual eligibility, linked account and remaining benefit rather than treating a ChatGPT plan as a general-purpose API balance. The official API pricing page explains usage charges, and the current account's billing settings establish what funds your requests.
The same principle applies when comparing product features. Access to a model through an API does not necessarily reproduce the provider's entire consumer application, including its interface, integrations or workflow-specific tools.
That distinction matters when someone says they replaced a subscription with “the same model.” They may have replaced one kind of inference while losing features they still need elsewhere.
Start with a workload, not a plan price
Choose a representative week or month of real work. For each recurring task, record:
- The type of output and what makes it acceptable.
- The model and settings actually used.
- Input and output usage when available.
- Paid searches, file processing, media and other tools.
- Attempts, revisions and final accepted outputs.
- The application features needed to review and deliver the work.
An isolated short chat is a poor proxy for a long document workflow. A small image preview is a poor proxy for several high-quality image revisions and a video clip.
Measure the workload at the level you buy. If your business needs ten usable product assets, the denominator is ten accepted assets, not the total number of generations. Rejected outputs still matter to the cost calculation.
Calculate model usage carefully
For a simple token-priced text request, the starting formula is:
Text model cost =
input tokens / 1,000,000 × input rate
+ output tokens / 1,000,000 × output rate
Use the actual rate structure of the selected model. Cached input, longer context, additional reasoning, service speed and tools may change the calculation. Do not apply an advertised cheapest rate to every kind of request.
Then calculate the complete route:
BYOK route total =
application fee
+ provider usage
+ separately billed tools and services
+ applicable taxes or other charges
Subscription route total =
subscription fee
+ paid usage beyond or outside included allowances
+ other required services
+ applicable taxes or other charges
If you are comparing spending already committed for another purpose, keep both views: incremental cost and total monthly cost. A subscription already paid for may make the next task inexpensive, but it still belongs in a full budget review.
Three hypothetical workload examples
Assume a fictional text model priced at $2 per million input tokens and $10 per million output tokens. Each task uses 4,000 input tokens and 1,000 output tokens. Assume a separate BYOK application fee of $12 per month.
These assumptions are deliberately simple. They exclude caching, tools, media, retries, tax and human review. They do not describe AI Magicx or a named provider's current plan.
| Workload | Tasks per month | Input tokens | Output tokens | Model usage | Application plus model usage |
|---|---|---|---|---|---|
| Occasional | 200 | 800,000 | 200,000 | $3.60 | $15.60 |
| Regular | 1,000 | 4,000,000 | 1,000,000 | $18.00 | $30.00 |
| Heavy | 4,000 | 16,000,000 | 4,000,000 | $72.00 | $84.00 |
For the occasional workload, input costs $1.60 and output costs $2.00, producing $3.60 in model usage. Adding the fictional application fee gives $15.60.
Now compare a fictional $30 subscription. The regular BYOK scenario also totals $30 before extras. But that does not make the two offers equivalent. You still need to know whether the subscription permits that workload, includes the same model and tools, and provides the features your team needs.
The heavy BYOK example shows why a low application fee is not a monthly spending ceiling. More work can increase provider charges even when the interface price stays the same.
Conversely, a subscription's advertised fee is not proof that a large workload is fully covered. Check its allowance, fair-use rules, expensive-model limits and paid top-ups.
Add media, retries and tools separately
Text token arithmetic does not predict image, video or speech costs. Those tasks may be priced by output, duration, characters, resolution or another usage unit.
Make a separate line for each task. For a product campaign, that might include copy drafting, image attempts, accepted high-resolution images, voice samples, final narration, video attempts and editing.
Also distinguish a failed technical job from an output you simply reject. Billing behavior depends on the service and the failure. Do not assume every unsuccessful creative attempt receives a refund.
Keep the amount billed and the acceptance decision as separate fields. That lets you see whether a more expensive generation route produces enough usable work to justify the difference.
Understand funding in AI Magicx
AI Magicx uses mana for paid work, with the estimate shown before the task starts. Supported own-key routes bill the relevant provider account for model usage. Platform-run work, such as search or sandbox use, can still use mana.
Using your own key does not upgrade your plan's feature access. For example, a key does not make a video tool available on a plan that does not include video. Check current plans and funding options for the actual work you want to do.
Before approving a task, identify which parts use your provider account and which parts use the workspace's balance. Compare those lines with the completed task record afterward. A zero model charge in one balance does not mean the provider did not bill another account.
Set limits before increasing usage
Use the controls each provider and application actually offers. Confirm whether a budget setting is a hard stop, a notification or a delayed report. Do not assume an alert prevents additional spending.
Keep keys in the application's supported key settings, never in a prompt, shared document or public example. Use separate provider projects or keys where appropriate for your team, and revoke credentials when they are no longer needed.
If a task stops for insufficient funds or access, keep the source material and diagnose the route. Check the application allowance, provider balance, model eligibility and key status separately. Repeatedly retrying the same request can create confusion without resolving the cause.
Choose the route that fits the work
BYOK can be attractive when you want provider-level usage records, already have a suitable provider account and prefer paying for a variable workload directly. It still requires attention to account billing and the application fee.
A subscription can be attractive when its included workflow and allowances fit your regular work and reduce the number of separate accounts you manage. It still needs an honest check of limits and additional charges.
A mixed approach may be reasonable: use included everyday work where it fits, and choose an own-key route for eligible tasks when you understand its costs. The right choice can differ by task rather than by person.
For help selecting the application around that choice, read the all-in-one AI platform guide. Use the cost worksheet to compare one representative month, then review the actual bill before expanding your workload.