Guide
API pricing vs a monthly subscription
We make a client that uses your own API keys, so we benefit if you conclude that the API is cheaper. Read this with that in mind — and note that the honest answer below includes the case where a subscription wins, and the case where neither is the right answer. The arithmetic is simple enough to check by hand, and every number in the calculator is one you can change.
The short answer
At mid-tier model pricing, with short conversations, you need roughly sixty messages a day, every day, before pay-per-token billing costs more than a twenty-dollar monthly subscription. Most people send nowhere near that, which is why the API is usually cheaper for individuals — often by a factor of five or ten.
But volume is not the lever people expect. Conversation length matters more than message count. Because the entire history is re-sent with every turn, a habit of long threads can drag the break-even from sixty messages a day down to about twenty. Someone who sends thirty messages a day in long conversations can easily spend more than someone who sends eighty in short ones.
Work it out with your own numbers
Change any field. The prices are pre-filled with round figures in the region of mid-tier model pricing as of September 2026 — they are a starting point, not a quote, and providers change them. Put your provider's current numbers in.
Includes the conversation history re-sent each turn
The calculator needs JavaScript. The worked examples below are the same arithmetic, done in advance.
Worked examples
All rows assume 500 output tokens per message and prices of $3.00 and $15.00 per million input and output tokens, against a $20 subscription. Only usage changes.
| Pattern | Messages/day | Input tokens | Per month | vs $20 |
|---|---|---|---|---|
| Occasional, short threads | 5 | 1,000 | $1.57 | Subscription costs 13× more |
| Daily use, short threads | 20 | 1,000 | $6.30 | Subscription costs 3× more |
| Heavy use, short threads | 50 | 1,000 | $15.75 | API still cheaper |
| Break-even, short threads | 63 | 1,000 | $19.84 | Line ball |
| Break-even, long threads | 21 | 8,000 | $19.84 | Line ball |
| Daily use, small/fast model | 50 | 1,000 | $1.31 | Break-even is ~762/day |
The last two rows are the ones worth remembering. Moving from short threads to long ones cuts the break-even by two-thirds. Moving from a mid-tier model to a small fast one raises it more than tenfold — at $0.25 and $1.25 per million tokens, you would need to send around seven hundred and sixty messages a day to reach twenty dollars.
What the arithmetic leaves out
A cost comparison that only counts tokens is not honest about either side. In the subscription's favour:
- A flat price never surprises you. Per-token billing is variable by design, and a month where you lean on it hard is a month you pay for
- Heavy users get a genuine bargain. Above the break-even the subscription is not merely convenient, it is cheaper, and it keeps getting cheaper the more you use it
- Subscriptions include things the API does not expose — the vendor's own app, and features that never appear as an API parameter. These differ by vendor and change often, which is exactly why this page does not try to enumerate them
- A subscription is one decision. Keys, balances, and per-provider billing consoles are administrative work that a monthly plan does not ask of you
In the API's favour, beyond the price:
- An idle month costs nothing. Usage that comes in bursts is billed as bursts; a subscription charges through the quiet weeks
- One key, many models. Keys let you put a frontier model, a cheap fast one, and a local one behind the same window and pick per task — a subscription buys one vendor's models
- No seat to cancel. Stopping is not an act of admin
The third option the comparison usually omits
A local model costs nothing per token, because it runs on hardware you already own. If your work is mostly summarising, drafting, rewriting, or answering questions about text you supply, a recent open-weights model on a decent laptop does it acceptably and the marginal cost is electricity.
This is not a claim that local models match frontier ones — they do not, and for hard reasoning or long-context work the difference is obvious. But the comparison is rarely all-or-nothing. The cheapest arrangement for many people is a local model for the routine bulk and a paid key for the work that genuinely needs a frontier model, which costs a fraction of either plan on its own. Conduit connects to Ollama and LM Studio alongside the hosted providers, so switching between them is a dropdown rather than a different application.
How to get your real numbers
Every estimate on this page is worth less than one week of your own measured usage. Two ways to get it:
- Your provider's billing console reports tokens per day, broken down by model. This is the authoritative source — it is what you will actually be invoiced for
- A client that tracks cost locally. Conduit keeps a per-day chart and token totals computed on your own machine. It labels them estimates, deliberately: the pricing table behind them is a small hand-maintained subset, and the provider's invoice is the authority
Run whichever you pick for a week, multiply by four, and put the result into the calculator above. If the answer is close either way, take the subscription — the certainty is worth the few dollars. If it is not close, it usually is not close at all.
Questions
Is the API cheaper than a $20 subscription?
For most people, yes — by a wide margin, because most people send far fewer messages than they think. With mid-tier model pricing and short conversations, it takes roughly sixty messages a day, every day, before per-token billing reaches twenty dollars a month. But the number collapses as conversations get longer: the whole history is re-sent with every turn, so a habit of long threads can push the break-even below twenty-five messages a day. Your own usage is the only number that settles it.
Why does a long conversation cost so much more?
Because models have no memory between turns. To answer your tenth message the model has to be sent the previous nine along with it, so input tokens grow with every exchange while output stays roughly constant. A short exchange might send a thousand tokens of context; the same conversation twenty turns in can send eight thousand or more for an answer of the same length. Starting a fresh conversation when the topic changes is the single cheapest habit available to you.
Does a Claude or ChatGPT subscription include API access?
No. They are separate commercial relationships and separate credentials. A subscription pays for use of the vendor's own app at a flat rate; API access is billed per token against a key you create, with its own balance. Paying for one does not give you the other, and a client that uses API keys cannot use your subscription. That is worth knowing before you cancel anything.
What does a subscription buy that the API does not?
A price that never surprises you, which is worth more than it sounds — heavy users on a flat plan are getting a genuine bargain, and no amount of arithmetic changes that. It also buys the vendor's own polished app, and features that are not exposed through the API at all, which differ by vendor and change often. If you use one model heavily and want none of the flexibility, the subscription is simply the better product.
How do I find out what I actually use?
Measure rather than estimate — the guesses people make about their own volume are usually wrong by a factor of several. Every provider shows per-day token usage in its billing console. If you use a client that tracks cost locally, that works too: Conduit keeps a per-day chart and token totals computed on your own machine, though it labels them estimates, because the pricing table behind them is a hand-maintained subset and the provider's invoice is the authority.
This page deliberately names no vendor's subscription price or token rates, because both change and a page asserting them goes stale silently. The pre-filled figures are round numbers in the region of mid-tier model pricing in September 2026, provided so the calculator does something useful before you touch it — check your provider's own pricing page for what you would actually pay. Conduit's own behaviour is checkable against the documentation and the source. Found an error? Tell us and we will fix it.
Related
- How to demo an MCP server to a customer — Hosting an MCP server is not a customer demo. Compare a hosted chat, Claude Desktop with an extension, and your own client — and what you can ship today.