The spiral has a specific anatomy: ChatGPT Business seats ($20/user annual) carry Plus-level Codex windows per person, so daily-driver devs blow through them, overflow goes to credits billed per model (GPT-6 Astra at ~50x GPT-5.6 Luna rates), and anything running on raw API keys is unmetered by anything except the invoice. Fixing the number means separating the two workloads: interactive Codex stays on seats — that part is genuinely fixed — and the volume work (long agentic runs, background tasks) moves to a custom provider with a hard budget. Codex CLI supports that natively via [model_providers] in config.toml, which is how a team of five gets one predictable line item instead of five surprise ones.
| Component | Billing shape | Predictability |
|---|---|---|
| ChatGPT Business seats | $20/user/mo (annual) | Fixed — but windows are Plus-level, sized for interactive use |
| Credit overages | Per-model rates, top models ~50x the light ones | The spiral: bought when devs are blocked, priced at panic rates |
| Raw API keys for agents | Pure metered | None — this is usually the line that doubles |
| Flat routed budget for volume work | $19–249/dev-group fixed | Exact — and it absorbs precisely the workload that causes overages |
Trying to cover all Codex usage with one billing mode is the mistake. Seats are excellent value for the interactive layer — reviews, quick edits, questions — because a human at a keyboard rarely outruns the window. Agentic volume is the opposite: one long refactor session consumes dozens of requests, and it is exactly this traffic that turns into credit purchases and API spikes.
Route by workload instead: seats for humans, a budgeted endpoint for agents. Codex CLI makes the split trivial — profiles let a dev flip between plan sign-in and the custom provider per invocation, so the same tool bills to the right bucket depending on what it is doing.
The fixed-number setup caps spend, not appetite: if the team's agentic volume outgrows the flat budget, the work paces down or you buy the next tier — a visible, planned decision. That beats the alternative, where the same growth shows up as an overage invoice nobody approved.
Standard Compute is the budgeted half of the split: Codex CLI points at api.stdcmpt.com/v1 (both /v1/responses and /v1/chat/completions), business plans give the team one subscription with per-dev budget slices ($499/mo for 5 seats), and smart routing sends routine agent turns to efficient models so the budget survives daily use. About a quarter of our active customers run Codex CLI this way.
Business seats carry Plus-level allowances per user — the value is per-seat management and admin controls, not bigger windows. Teams whose devs cap out on Plus will cap out on Business too; that workload needs credits, Pro seats, or an external budget.
No — plan allowances are per account, and sharing accounts violates terms. Pooling happens legitimately at the API layer: a shared endpoint with per-dev keys drawing on one budget, which is what flat team plans implement.
Real-world band: $100/mo for interactive seats, plus $150–250 per agent-heavy developer if metered — so $500–1,300 total. A flat setup compresses the agentic half to a fixed $499-tier number; teams above ~10 devs step to the $999 tier.