THE VERDICT
● Subscription if your usage is human-paced and you want the bill to be the same every month.
● API if anything runs unattended, or if usage varies by more than about half month to month.
● The tell: if a loop can run while you sleep, a weekly pool is the wrong shape.
What changed this month
| Vendor | Change |
|---|
| Anthropic, today | Claude Code weekly limits settle below the promotional level |
| Google, 2 September | Gemini Notebook moved to compute-specific flexible limits |
| OpenAI, late August | Restored a strict five-hour cap on Codex and Work tools |
THREE VENDORS, ONE FORTNIGHT, SAME DIRECTION
That is not coordination. It is three companies reaching the same conclusion about the same workload — agents consume far more than chat, and flat-rate pricing was set before anyone knew what heavy use looked like.
Assume the direction continues. Every change this quarter tightened rather than loosened.
The two shapes
| Subscription | API |
|---|
| Bill | Same every month | Varies with usage |
| Ceiling | Yes, and it moves | None |
| Terms change | Vendor decides, as today showed | Rate card, published |
| Unattended work | Badly suited | Suited, with a spend cap |
| Worst failure | You stop working mid-task | A loop bills all night |
The question that decides it
Not how much you use, but whether anything runs without you watching.
Human-paced work fits a subscription well — you cannot exhaust a weekly pool by typing. An agent loop can exhaust it in an afternoon, and it can also run up an unbounded API bill overnight. Those are the two failure modes, and they are opposites.
So the real answer for most teams is both: a subscription for interactive work, and API with a per-run spend cap for anything automated.
Which one
| If you... | Pick |
|---|
| Type at a keyboard all day | Subscription. You will not exhaust it |
| Run anything on a schedule | API, with a per-run cap set before you start |
| Need the bill to be predictable | Subscription, and accept that the allowance can move |
| Have wildly variable months | API. You stop paying for the quiet ones |
| Do both kinds of work | Both. Most teams end up here |
FAQ
Is the API cheaper than a subscription?
Depends entirely on volume. At Sonnet 5 rates, $100 buys roughly 33 million input tokens. Price a month of your actual usage rather than guessing.
Will limits keep tightening?
Every change this quarter went one way. Plan on the assumption that they do, rather than that a promotion returns.
What is the main risk with API billing?
An unattended loop running overnight. Set a per-run spend cap rather than a monthly one — a monthly cap can be exhausted in an afternoon.