# What a 24/7 AI Agent Actually Costs: API vs Subscription

A 30-day worked scenario at three volumes: $239 to $7,155 a month on metered GPT-5 pricing, against a flat $129 fee plus a $20 to $200 ChatGPT plan. Estimates, math shown.

*Published 2026-06-12 · https://proxyllm.ai/blog/what-a-24-7-ai-agent-costs*

An always-on agent costs between roughly $240 and $7,200 a month at June 2026 GPT-5 API rates, depending on volume. The same three workloads run flat at $129 a month plus a $20 to $200 ChatGPT plan through a subscription-backed endpoint, using capacity estimates rather than guarantees. Here is the 30-day arithmetic at three volumes, both ways.

## The reference agent

To make the numbers honest, fix the shape first. Our reference 24/7 agent (monitoring, triage, scheduled jobs, event-driven tasks) averages:

- 12 model calls per task
- 5,000 input / 700 output tokens per call, so 60,000 input and 8,400 output tokens per task
- GPT-5 at $1.25 input / $10 output per million tokens (June 2026, [live prices at openai.com/api/pricing](https://openai.com/api/pricing))

Per task: 0.06 x $1.25 + 0.0084 x $10 = $0.075 + $0.084 = $0.159. If your agent is shaped differently, the formula to rebuild this table is in [how to calculate AI agent costs](/blog/ai-agent-cost-calculation).

## The 30-day API arithmetic

| Volume | Tasks/day | Monthly tokens (in / out) | GPT-5 API cost |
| ------ | --------- | ------------------------- | -------------- |
| Light  | 50        | 90M / 12.6M               | $238.50        |
| Medium | 300       | 540M / 75.6M              | $1,431         |
| Heavy  | 1,500     | 2.7B / 378M               | $7,155         |

Checking the medium row by hand: 9,000 tasks x 60k input = 540M input tokens, which is 540 x $1.25 = $675. Output is 9,000 x 8.4k = 75.6M tokens, which is 75.6 x $10 = $756. Total $1,431 a month, every month, for as long as the agent runs.

Model choice moves the whole table: on GPT-5.5 ($5/$30) multiply by roughly 3.5x, on GPT-5 Mini ($0.25/$2) divide by 5. Prompt caching and batch discounts shave the metered number but never change its shape: it scales with every task the agent takes on.

## The subscription math

ChatGPT plans include Codex, and [Codex Hosted](/) runs OpenAI's official Codex CLI signed into your own account, exposed as an OpenAI-compatible endpoint. The workload then bills against the flat plan instead of the meter. Our planning estimates for plan capacity: Plus ($20) absorbs roughly $700 of API-equivalent work a month, Pro 5x ($100) roughly $3,500, Pro 20x ($200) roughly $14,000. These are estimates from observed usage windows; OpenAI adjusts limits over time.

| Volume | API cost | Covering setup     | Flat cost          | Monthly difference |
| ------ | -------- | ------------------ | ------------------ | ------------------ |
| Light  | $238.50  | Plus + ProxyLLM    | $20 + $129         | ~$90               |
| Medium | $1,431   | Pro 5x + ProxyLLM  | $100 + $129        | ~$1,200            |
| Heavy  | $7,155   | Pro 20x + ProxyLLM | $200 + $129        | ~$6,800            |

At medium volume, the same 24/7 agent costs $1,431 on the meter and $100 for a Pro plan plus $129 for us. The $129 is our flat SaaS fee; we add no inference markup, so the flat cost does not grow with task count. Five more scenarios with full tables are in [Codex Hosted savings examples](/blog/codex-hosted-savings-examples).

## Why always-on agents suit usage windows

A 24/7 agent is the friendliest workload a usage window will ever see: steady draw around the clock, no burst that slams a window in an afternoon. Bursty workloads (a launch-day backfill, a batch migration) are the ones that exhaust windows early.

The honest caveats, so the flat number stays trustworthy:

- **Windows, not budgets.** Capacity arrives in rolling windows. A heavy day can exhaust one early; requests then fall back to a second connected account, then to your own API key until the reset. The mechanics are in [what happens when you hit the Codex usage limit](/blog/what-happens-when-you-hit-codex-usage-limit).
- **Complete responses on the Codex lane.** No token streaming; the response arrives whole. Background agents do not care. A live chat UI does, so keep that on the API lane.
- **Estimates are estimates.** Size against the middle of the capacity range. If your metered bill sits near a tier's ceiling, plan for the next tier or a second account ($100 steps, not linear growth).
- **Programmatic Codex use is documented OpenAI functionality, but OpenAI has the final call** on its accounts and services.

## What moves your number

Four inputs dominate: tasks per day (linear), calls per task (linear, and the easiest to trim), model choice (5x to 20x swings), and retry rate (a silent 10 to 25% tax). Get those four from a week of logs and the 30-day projection is usually within 15%.

Put your own volumes into the [calculator](/calculator); it runs this table for your numbers and picks the covering tier, including multi-account setups past Pro 20x.