Subscription Seats vs API Keys: How a Small Team Decides

Maker's note: 기준일 2026-10-03. The general framework below is opinion, offered as such. The AZET-specific parts stay inside what the project README documents — including its explicit non-claims about live paid operation.

The question every small team hits

Sooner or later, a team using AI assistants daily faces the same fork: pay monthly seats for the subscription products, or hold API keys and pay per token. Both are legitimate answers. The mistake isn't picking either one — it's picking by vibe, then being surprised by the bill or by the ceiling. Here's the framework we use, and where AZET sits in it.

The general tradeoff

Subscription seats buy predictability. A flat monthly price, a polished app included, and per-seat limits that are easy to reason about. The failure mode is waste: seats that sit idle most of the month still cost the same, and heavy parallel workloads can bump into what a seat allows.

API keys buy proportionality. You pay for what you actually run, which suits bursty and scripted agent work, and metered access composes cleanly with automation. The failure mode is the opposite of waste: a runaway loop with no cap, or a quiet month followed by a shocking invoice, because nobody was reading the meter.

Dimension	Subscription seats	API keys
Cost shape	Flat, predictable monthly	Metered, usage-sensitive
Best workload	Interactive daily use by a person	Bursty, parallel, scripted agent work
Typical failure	Paying for seats that sit idle	A runaway loop with no cap
Governance	Seat count, nothing more	Budgets, caps, and a ledger someone reads

Four heuristics cover most small-team cases. First, utilization: a person who uses an assistant all day almost justifies a seat; a person who touches it weekly does not. Second, shape: interactive work fits seats, batch and parallel work fits meters. Third, attention: if nobody will watch usage, prefer the flat price until someone will — an unwatched meter is a liability, not a feature. Fourth, measure before optimizing: neither seat math nor token math survives contact with guesses.

Where AZET sits: bring your own CLI

AZET's stance is that your existing subscriptions stay yours. The app connects to the official Claude and Codex CLIs you already use: azet login --provider claude or azet login --provider codex verifies the authentication state and models of an official CLI you have already logged into. No credentials are copied, the provider login and logout are never performed on your behalf, and logging out of AZET does not revoke the provider's OAuth session. New provider authentication happens in the official CLI itself.

The cost accounting follows that stance. Usage through an official subscription CLI is recorded in AZET's ledger as 0 hosted credits — a deliberate distinction, not a free lunch: AZET does not measure your provider's subscription billing at all. Your seat, your bill, your existing terms. For a small team, this means adopting another tool doesn't force a procurement decision about model spend.

The hosted side, documented honestly

AZET also has a hosted-credit design, and because we write about costs the way we ask you to reason about them, here is what the README actually says. One credit equals $0.01. The Free tier is 500 credits per UTC month. Pro is $15 per valid payment-invoice cycle with 2,000 credits included. Team is designed at $40 per seat, but it is not activated because the included credits are still undecided. Hosted text generation runs a fixed registered model (azet/gpt-4.1-mini-2025-04-14) with token pre-counts, maximum-cost reservations, actual-usage receipts, and a 1.3x multiplier over measured cost.

Two boundaries on those numbers. First, live paid operation — real provider keys and real payment flows — is a separate completion condition the project explicitly does not claim as done, so treat the tiers as documented design, not a live price list. Second, this describes the AZET app's hosted ledger. The azet.io portal assistant's pricing remains unpublished as of October 2026, as we stated in our introduction post. Different surfaces, both stated plainly.

One detail worth stealing for your own stack: the ledger distinguishes azet credits (local, per-task) from azet credits --cloud (the authenticated server ledger) as two different datasets, and it reconciles pending requests from preserved server-signed receipts without re-calling the model. When a cost number can be replayed from a receipt instead of re-measured, trust it more.

The actual decision

For most small teams the answer is hybrid: seats for the people who live in the product all day, keys for the scripts and agents that run at night. And for tools like AZET that let you bring the subscriptions you already pay for, the fork stops being either-or — your seats keep their value, and the hosted meter exists only for when you want it.

If you want to watch the azet.io portal take shape, the waitlist at https://azet.io is open.

FAQ

Is one side simply cheaper — seats or API keys? Neither, in the abstract. Seats waste money when they sit idle; unwatched meters produce shocking invoices. Utilization, workload shape, and whether anyone reads the meter decide it — and neither kind of math survives contact with guesses.

Does using AZET with my existing subscription cost extra? Usage through an official subscription CLI is recorded in AZET's ledger as 0 hosted credits, and AZET does not measure your provider's subscription billing at all. Your seat, your bill, your existing terms.

Are AZET's hosted tiers a live price list? No — treat them as documented design. Live paid operation, with real provider keys and real payment flows, is a completion condition the project explicitly does not claim as done, and the Team tier is not activated because its included credits are still undecided.

What's the difference between azet credits and azet credits --cloud? They're two different datasets: the local, per-task ledger and the authenticated server ledger. The cloud side reconciles pending requests from preserved server-signed receipts without re-calling the model.