A capacity problem has two exits: write the code that makes the loop cheaper, or buy the plan that makes the loop legal. Both cost something, and teams pick by instinct more often than by arithmetic. The honest comparison prices the engineering week the same way it prices the plan.
Put both exits in money
The plan side is already in money. The jump from REST snapshots at $99 a month to high-volume REST at $229 is $130 a month, and from the combined REST + SSE plan at $149 it is $80, as the plan guide lists the terms, checked 2026-10-29.
The code side needs converting. A week of engineering on the poller is not free; it is a week of someone's salary plus everything else that did not ship. Price it at whatever a week costs your team, then ask how many months of plan delta that week equals. At almost any real salary, one optimization week costs more than a year of the upgrade. The plan is the cheap exit far more often than instinct admits.
What the plan buys without a deploy
The upgrade arrives the day you pay for it. No branch, no review, no regression risk, no rollout. What it buys is headroom: the room for retries, bursts and match-day churn above your steady rate, which the pricing page is explicit about needing, since a per-second limit is not permission to run permanently at the ceiling, checked 2026-10-29.
Headroom bought with money also unbuys itself the month you downgrade, which is a feature when the busy season is a season.
What the code buys that the plan cannot
Optimization compounds differently. A poller rebuilt around the cursor, caching and even pacing costs less to run forever, on every plan, at every scale. The work survives plan changes, provider changes and growth. If the loop is wrong, wasteful by design, the bigger plan is a subscription to your own bug, paid monthly.
That is the asymmetry that matters: the plan rents capacity, the code removes the need for it.
The decision rule
Buy when the loop is already right. A well-built pipeline brushing its ceiling on busy days has an honest load, and headroom is a purchase, not a project. A 429-driven emergency is always a buy first, because the fix ships at the speed of a checkout.
Optimize when the bill says the loop is wrong: full snapshots in steady state, per-user requests reaching upstream, bursts at the top of every cycle. Those are defects, and no plan size cures a defect for more than a month at a time.
Price the week, price the plan, and notice which one keeps charging you. The cheap answer is usually the plan today and the code eventually, in that order.