3,119 AI Agent Sessions on a Flat Subscription (and the $58 Afternoon)

I counted my agents. Not the names, the sessions: every Claude Code and Codex run still sitting on my two laptops. 3,119.

Together they read 38 billion tokens and wrote 120 million. My agents spend 99.7% of their lives reading and 0.3% saying anything new.

On a pay-per-token API, that habit is a five-figure invoice. On two flat subscriptions, it's the same bill as any other month. One afternoon I learned the difference the expensive way.

What does a fleet of AI agents cost on a flat subscription?

The same fixed monthly fee, whether the agents work one hour or all night. You swap an invoice for usage limits, and a fleet is exactly the workload where that swap pays.

Here's my fleet, priced as if metered. The Claude half covers only the five weeks of logs still on disk: 21.6 billion cached tokens, 365 million cache writes, 53 million tokens of output. At Anthropic's published Opus 5.5 rates, that's about $7,200, and that's a floor, since some turns ran on pricier models.

The Codex half re-read 15.7 billion cached tokens. With no verified public price for those models, I used the rate OpenAI actually charged me on my one bad afternoon: 58 cents per million cached tokens. Another $9,000 or so, for cached input alone. Even at Claude's cheaper cache rate it would be about $3,100.

So: somewhere between $10,000 and $16,000 of agent work, on two flat subscriptions.

The catch: limits. A subscription is a weekly allowance, and a fleet can eat it. So a small guard watches the fuel gauge. It polls Claude's weekly usage, reads the percentage Codex logs about itself, and kills only the lanes (one agent each) of whichever engine crosses my cap. Worst case, the agents take a nap.

That's the whole rig behind my overnight agents: fifteen lanes for eleven hours, no meter running.

Is Devin or Replit worth it compared to your own agents?

For me, no, and I've used both properly. To me they feel like products for people who don't want a terminal. I want the terminal. My rig is faster and more honest, and it costs the same every month.

Devin. One client had Devin wired into its pull requests. It wrote "Comprehensive QA Review" comments, always felt laggy to me, and its confidence outran its results. My favourite afternoon, straight from the notification emails:

1. 14:52Devin opens a PR for a small filter feature
2. 14:56PR closed
3. 15:00Opens it again. "Closing - incorrect target branch (staging instead of master)."
4. 15:04Again. "Closing - still targeting staging instead of master."
5. 15:08Fourth try, closed 13 seconds after opening: "still targeting staging despite explicit master base_branch parameter."

Four pull requests, sixteen minutes, one wrong branch. My agents blunder too, but they ship through one push script and answer to a separate tester, so the blunder dies in a terminal instead of in a stream of notification emails. More on that in why every agent needs a QA department.

Replit. I've used it a lot; it has sent me 116 publish notifications. It's genuinely good at the prototype you later port to a real codebase.

But I've also typed, typo fixed, "Replit cannot seem to set secrets himself". And the billing is the opposite of mine; paid agent work uses "effort-based pricing based on request complexity". One workspace mailed me "has used 100% of Credits" twice.

Metered agents put a meter on your curiosity. Every "try it again" costs money, so you stop asking. On a flat fee, I run three independent QA reviews on one fix, and nobody sends a receipt.

Replit sells an agent that can work for 200 minutes. Mine work twenty minutes or all night, and the fee doesn't move.