Every AI tool vendor has to answer the same billing question: subscription tiers or prepaid credits? As someone who buys from roughly 30 providers and sells through one product, I have strong opinions formed from both sides of the invoice. The short version: subscriptions were designed for humans, agents are not humans, and prepaid credits are the model that actually matches how agent workloads spend money. But the argument is worth making properly, because the subscription model’s failure modes are subtle until they are expensive.
What subscriptions assume
A subscription tier (“$49/month, 5,000 requests included”) is a bet on predictability. It works beautifully when usage is smooth and human-shaped: a person uses a product roughly as much this month as last month, the vendor gets forecastable revenue, the buyer gets a forecastable bill. Fine.
The model embeds two assumptions. First, that you can predict next month’s usage well enough to pick a tier. Second, that usage fills the tier steadily, so the included quota is neither mostly wasted nor suddenly exhausted. Human usage often satisfies both. Agent usage reliably violates both.
Agents are spiky, and tiers punish spikes
Agent workloads follow the work, not the calendar. You run a backfill and generate a month of transcription volume in an afternoon. A product launch triples search traffic for a week, then it subsides. A pipeline is dormant for three weeks, then reprocesses everything. Under tiered billing, this shape forces a bad choice at every billing cycle: size the tier for the spike and waste the quota in quiet months, or size it for the baseline and hit overage rates (or hard caps, mid-backfill) whenever real work happens.
Prepaid credits dissolve the choice, because spend simply follows usage. Quiet month, credits sit there. Spike, credits drain in proportion to the work done, at the same unit price. There is no tier to outgrow and no quota to waste, and the unit prices stay legible: a search costs what a search costs, whether you ran ten this month or ten thousand. That per-unit legibility is something I have written about before in understanding per-unit pricing, and it only exists cleanly under usage-based billing.
The runaway-agent problem
Here is the argument I consider decisive, and it is about failure, not convenience.
Agents are automation with money attached. A human with a subscription cannot spend faster than they can click. An agent in a malformed loop (a retry cycle that never terminates, a planner that keeps deciding it needs one more search, a batch job pointed at the wrong dataset) can fire paid API calls as fast as the network allows, around the clock, with nobody watching. Every agent builder eventually writes this bug. I have written it.
Now compare the blast radius under the two billing models. On post-paid, usage-metered billing (subscription with overages, or pure pay-as-you-go against a card), the runaway loop runs until someone notices, and the bill is whatever it is. The horror stories in this genre are all post-paid stories. On prepaid, the loop runs until the balance hits zero, and then every call fails. Your worst case is a number you chose in advance, because funding the balance is setting the cap.
That is the deep fit between prepaid and agents: prepaid turns the billing model itself into a spend cap. Not a soft alert, not a dashboard you should have checked. A hard financial boundary that a buggy loop cannot cross. For autonomous systems, “the default failure mode is bounded” is worth more than any convenience feature, and it is the main reason route.tools is prepaid: you top up a balance, usage draws it down, and a runaway agent exhausts a balance instead of a credit card.
What honest prepaid looks like
Prepaid is not automatically virtuous; vendors can abuse it too (expiring credits, opaque burn rates, units that obscure real prices). The version I consider honest has a few properties, and they double as a checklist for evaluating any credits-based vendor:
- Legible unit prices. You should know what one call costs before making it. Our prices are provider list plus 20%, stated publicly in the pricing docs, so the markup is a number, not a mystery.
- Per-call receipts. Every response through our router includes the exact price of that call, so reconciling “where did my balance go” is grep, not archaeology.
- A cheap on-ramp. Prepaid’s advantage is trying before committing; signup comes with $2 of free credits, which at, say, $1.20 per 1k routed search queries is a genuinely meaningful test budget rather than a token gesture.
- No tier games. No feature gating by spend level, no expiring balances pressuring you to use it or lose it.
Where subscriptions still make sense
Fairness requires the other side. If your usage genuinely is flat and predictable, a well-sized subscription can beat list-price usage billing, since vendors discount the predictability. And negotiated enterprise commitments are effectively subscriptions with better math. There is also a hybrid worth knowing about: if you already hold direct provider subscriptions or negotiated rates, BYOK lets you route with your own keys at a 5% platform fee, keeping your subscription economics under a routed interface.
But for the default case (agent workloads, spiky by nature, capable of spending money autonomously) prepaid is the model whose failure mode is a stalled agent rather than a surprise invoice. If you want to see the model in practice, the dashboard starts you with the $2 in credits and a running per-call ledger.