A budget warning is not a spending limit

Shared reading from the original source. Read the piece, then bring your experience or a different take.

S
Reading from

Simon Willison

Simon Willison’s Weblog · 2026-10-03

A warning email does not stop the next API call.

That is the practical issue in Simon Willison’s October 3 argument for default hard budget caps. Coding and personal agents lower the effort needed to deploy something that consumes paid services. A notification can arrive while that service keeps running.

The uncomfortable tradeoff is availability. A hard cap can interrupt a useful job; a soft alert can preserve availability while allowing a larger bill. Neither choice should be hidden from the person paying.

For an agent pilot, ask what the limit covers. Model tokens, search, browser sessions and hosted compute may have separate meters. A limit on one does not establish a ceiling on the whole workflow.

One question to bring to a vendor demo: show what the next request does after the limit is reached. Does it fail, pause, switch service or continue? Do retries spend more?

Where would you draw the line: stop immediately, finish the current step, or ask for a one-time extension?

Read the originalWe’re going to need default hard budget caps on pretty much everythingSimon Willison’s Weblog ↗

From the public discussion

The disagreement is about production availability versus financial exposure. The replies then separate a spending ceiling from a complete service shutdown.

Selected comments from Hacker News. Each excerpt or summary links to its original.

  1. M

    motionlessveloc

    motionlessveloc reports support problems caused by hard cutoffs during growth spikes. They favor alerts for human-operated production services, where billing protection can cause costly outages.

  2. K

    KronisLV

    In reply to motionlessveloc

    Customers have different risk tolerances. Let each choose visible alerts or a hard stop.

    Want an alert? Configure one.

    Want a hard limit? Configure one.

  3. E

    edoceo

    In reply to KronisLV

    A cap chosen during setup can become wrong as a service grows. Revisit it during launch reviews and load checks; configuration needs a maintenance process.

  4. B

    Breza

    In reply to KronisLV

    The workload matters: a training VM should stop before an unattended weekend becomes expensive; a growing startup may prefer alerts while fixing inefficiency.

  5. M

    mybrowsercache

    In reply to motionlessveloc

    A spending ceiling could preserve limited service, as mobile plans do after a data allowance. Otherwise, this user prefers failure over unbounded charges.

  6. C

    collingreen

    In reply to mybrowsercache

    That changes what “hard” describes: spending can stop while the service continues in a degraded mode.

Have a result, correction, or a different constraint?

Start a related discussion