With BYOK you pay for inference
at cost, not at someone else’s markup.
The idea is simple: instead of buying credits from a tool, you bring your own provider key and the inference bill comes to you, at list price. What almost nobody explains is when that pays off and when it does not — because it does not always.
What each pricing model actually buys
BYOK does not remove the cost of AI. It moves it.
An AI tool charges you for two different things, even when it presents them as one price: the software — the interface, the orchestration, whatever they built — and the inference, which is what it costs for a model to read your text and write a reply.
Under the credit model you pay both to the same company, and that company pays the model provider. The gap between what they charge you and what they pay is their margin, and they usually do not publish it.
Under BYOK you only pay them for the software. You contract inference directly with Anthropic, OpenAI, Google or whoever you pick, and they bill you at the same public rate as anyone else. The tool routes your requests with your key and never touches that side.
Put differently: BYOK is not a discount, it is removing a middleman. If you use little, that middleman may be subsidising you; if you use a lot, they are overcharging you. Where those two curves cross is the whole decision.
Numbers, not adjectives
What a million tokens costs today
Public rates from each provider, read off their own pricing page on 17 August 2026. These numbers move; if you are going to decide with them, check the source first.
| Model | Input / 1M | Output / 1M | What you would use it for |
|---|---|---|---|
| Claude Opus 5 | $5 | $25 | Long agentic work and hard code |
| Claude Sonnet 5 | $2 | $10 | The balance for day-to-day work |
| Claude Haiku 4.5 | $1 | $5 | Short tasks and classification |
| GPT-5.6 Sol | $5 | $30 | OpenAI’s top tier |
| GPT-5.6 Terra | $2 | $12 | OpenAI’s mid tier |
| GPT-5.6 Luna | $0.20 | $1.20 | High volume, low cost |
| Local model via Ollama | $0 | $0 | No token bill; you pay in hardware and in quality |
Output costs five to six times what input costs at nearly every provider, and in agentic work input dominates consumption: every step resends the accumulated context. That is why an estimate built on a short chat lands far below what an agent working for an hour actually spends.
The part nobody publishes
When BYOK saves you money and when it just adds work
It comes down to one variable: how much you consume. Everything else is secondary.
Heavy or spiky usage
If you leave agents working for hours, or you have heavy weeks and dead weeks, flat software price plus provider rate almost always wins. Credit plans are sized for the average user, and agentic use is not average.
You already pay for an agent subscription
If you already have Claude Code or Codex, that fee includes inference you are only half using. A tool that signs in with your existing session instead of asking for a new key adds no cost at all.
You want to pick the model per task
With your own key you can send cheap work to a small model and hard work to a big one. In a credit system the provider makes that call, and usually not in your favour.
Light, occasional use
A handful of questions a week fits comfortably inside the free tier of almost any tool. Adding a key, a card and one more billing console for that is work with no return.
You do not want to manage keys
A key is a credential: it has to be stored, rotated and watched for spend. If you are a team with nobody owning that, the margin the credit model charges buys something real.
You need spend predictable to the cent
With credits you know the monthly maximum exactly, because it is the fee. With BYOK you set the maximum yourself in the provider console — and if you forget, there is no maximum.
The landscape, not an ad
Ways to work with your own key
Lienox is one of several, and for plenty of people it is not the right one. These are the real categories and what each one solves.
| Route | Software cost | When it is the right one |
|---|---|---|
| The provider API directly, with your own scripts | Zero | You can code and your flow is fixed. Nothing beats this on cost, and nothing costs more to maintain. |
| Terminal agents (Claude Code, Codex, OpenCode) | Subscription or your key | You work in the terminal and one session at a time is enough. |
| AI editors (Cursor and similar) | Fee, BYOK on some | The work is writing code and it happens inside the editor. |
| Desktop orchestrators (Lienox and its category) | Flat fee | You need several agents at once, on a schedule, with the work happening on your own machine. |
| Local model via Ollama or llama.cpp | Zero, plus hardware | The requirement is that no token leaves your network. You pay for it in model quality. |
Lienox charges a flat fee for the software — there is a free tier, and the paid ones are on the pricing page — and never touches inference: it works with your OpenRouter or NVIDIA key, with the Claude Code, Codex, OpenCode or Kimi subscription you already pay for, or with a local model via Ollama. What you buy is the orchestration, not the tokens.
Frequently asked
BYOK, specifically
What does BYOK actually mean?
BYOK stands for "bring your own key". It means the tool you use does not sell you inference: you contract an API key directly with the model provider, and the tool uses it to make requests. The provider bills you at its public rate, and the tool charges separately for its software.
Is BYOK always cheaper?
No. It is cheaper when you consume a lot, because the middleman margin disappears. On light usage it usually comes out worse: the free and entry tiers of credit-based tools are sized for exactly that use and often subsidise it. The practical rule is that if your usage fits comfortably in an entry plan, BYOK only adds admin work.
How do I control spend if the key is mine?
In the provider console, and it has to happen before you start, not after. Anthropic, OpenAI and Google all let you set a monthly spend limit and threshold alerts. It is the step most people skip, and the only one separating an unexpected invoice from a quiet night — especially if you are leaving scheduled tasks running unattended.
Can I use BYOK with the subscription I already pay for?
Depends on the tool. Some accept only an API key; others also sign in with an existing subscription session, which avoids paying twice for the same inference. Lienox supports both: your own key with OpenRouter or NVIDIA, and subscription sign-in with Claude Code, Codex, OpenCode or Kimi with no key pasted at all.
Where is my key stored, and who can see it?
That is the right question, and the answer changes the risk entirely. If the tool is a cloud service, your key lives on their server and you have to trust how they store it. If it is a desktop app, it can stay on your disk and never leave — which is how it works in Lienox. Before pasting a key anywhere, ask where it is stored and from where it is used.
How much does an agent actually consume while working?
Far more than a chat, for a structural reason: every step of the agent resends the accumulated context, so input consumption grows with the length of the task instead of staying flat. An estimate built on a few questions falls short by a large factor. The only reliable number is your own: run your real task for a day with the provider meter in front of you.
What if I do not want to pay for inference at all?
Then what you want is a local model, not BYOK. With Ollama or llama.cpp serving an open-weights model on your own machine there is no token bill: the cost is hardware and electricity. You pay for it in quality, because the models that fit on a desktop machine are still behind the best API models, and on long tasks that gap compounds.
Try it with your own key
The free tier does not cut features and does not ask for a card. Install it, paste your key or sign in with the subscription you already pay for, and watch the provider meter while it works.