Your Agents Eat Your Subscription

Halfway through writing this, the second agent I use for review died mid-draft. Not a crash, not a billing failure — You've hit your usage limit. A $200-a-month subscription, exhausted at 7:21 in the morning, resetting at 10:15.

That is the entire article in one sentence. The agent you send off to work unattended is not spending some separate automation budget. It is eating the same meter you type against.

I spent this week rebuilding how AI tooling gets paid for on my machine — a client card, two vendors, a company plan, and a stack of assumptions that turned out to be wrong. Most of what I believed about how this billing works was wrong in a way that costs money. Here is how it actually works in August 2026.

What I Actually Run

A client gives me a card with a monthly AI budget. Everything below is bought on my own personal accounts and used on my own machine — that distinction turns out to matter more than I expected.

The credits line is the interesting one, because it exists purely to cover unattended work. I run a loop that picks up tickets and ships them end to end without me watching — worktrees, commits, CI, deploys. That loop runs claude -p, the headless mode. And headless is where the assumption I had been carrying quietly fell apart.

There Is No Automation Budget

The assumption almost everyone makes: interactive sessions come out of your subscription, and headless runs — CI jobs, unattended batches, agent loops — come out of something else. An API bucket. A separate allowance. Something.

They do not. On both vendors, the same meter runs.

On Claude, claude -p and the Agent SDK draw from your subscription's usage limits, exactly like typing in the terminal. Anthropic announced on 15 June 2026 that they would split SDK usage off with its own monthly credit — $20 for Pro, $100 for Max 5x, $200 for Max 20x — and then paused it. Their current wording is unambiguous: Agent SDK, claude -p, and third-party app usage still draw from your subscription's usage limits.

On Codex it is the same story. CLI, IDE and web share one allocation and the same five-hour windows. There is no CLI-only pool.

What “Included” Actually Means

Claude runs a five-hour session window plus a weekly cap that applies across all models. Anthropic does not publish the numbers. Max 5x is described as five times Pro per session, Max 20x as twenty times Pro. What that means in messages or tokens, you discover in Settings → Usage, or by hitting the wall.

Codex publishes actual figures. Per five-hour window:

The ranges are wide because a message is not a fixed unit — context length, reasoning depth and tool calls all move it. OpenAI puts the average at 5–40 credits per message. But a wide published range still beats no published range. When you are deciding whether a batch fits before the window resets, one vendor lets you estimate and the other does not.

What Overage Actually Costs

Both vendors price overage at API rates. OpenAI states it outright: buying credits is neither a discount nor a penalty compared with using an API key. Anthropic bills usage credits at standard API pricing. Nobody is giving you a deal for staying inside their subscription — overage is list price, always.

Codex quotes overage in credits at $0.04 each. Converted to dollars per million tokens, against Claude's published API rates:

Look at the top two frontier rows. Input is identical. Output on Opus 5 is $25 against Sol's $30 — about 17% cheaper. My instinct said the opposite, and I would have bet money on it. Codex feels more generous because the included plan is roomier and the limits are published. That is the plan, not the overage.

Do not over-read this. Two models do not consume the same tokens for the same task, Terra and Luna are dramatically cheaper if your work fits them, and Claude's prompt caching cuts input cost by up to 90% while batch processing halves it. The real lesson is narrower and more useful: read the rate card, do not reason from vibes about which vendor is cheaper at the margin.

The Meter Starts at Zero

Here is a trap that would have quietly blown my budget.

I bought credits, then went to set a monthly spend limit. The panel showed a healthy positive balance and, right above it, a usage bar reading $0.00 spent, 0% used.

The monthly spend limit does not count credits you already purchased that month. It governs what happens from that moment forward. So a cap of $X permits $X more on top of whatever you have already spent. Set it to your remaining headroom, not your total budget — otherwise you have authorised roughly double what you think.

Anthropic's help docs do not actually define whether that meter counts dollars charged or credits consumed. I checked both the Pro/Max and the Team articles; neither says. Which is its own lesson: when the documentation is ambiguous about a spending control, configure for the worse reading.

Why Auto-Reload Is Off

Both vendors offer automatic top-up. Balance drops below a floor, card gets charged, work continues. I turned it off and buy one tranche manually each month. Three reasons.

Predictability on someone else's card. This is a client budget with a hard monthly ceiling. An auto-reload that fires eleven times during a runaway loop is a conversation I do not want to have.

Discounts scale with tranche size. Anthropic advertises up to 30% off credit purchases, and the discount grows with the pack. A stream of small automatic top-ups almost certainly pays closer to list than one deliberate purchase does.

A hard stop is a feature. When the money runs out, the batch parks instead of quietly spending more. I would rather find a stopped loop in the morning than an unexplained charge.

The honest cost: a long unattended run can die at 3am with no way to resume until I top up by hand. That is the price of the ceiling, and I am paying it deliberately.

The Seat Math Nobody Runs

This is the part that changed an architectural decision, and it is the number most people never check.

When a company says “we'll give you a seat,” that sounds like the larger option. Run the arithmetic. Claude Team has two seat types, and Claude Code is included on both:

A Standard seat is roughly one sixteenth of the session capacity of a personal Max 20x. Even a Premium seat lands at about a third. For someone running agents all day, the company seat is not automatically the bigger option — depending on the tier it can be a fraction of a personal plan, at a comparable price point.

Which means the useful question, when someone offers you a seat, is not “do we have a plan?” It is two questions: which seat type, and what credit allocation. “We have a Team plan” is not an answer — the tier is the whole number.

On the corporate side, Team and seat-based Enterprise credits are configured org-wide by an owner, with a monthly spend limit and per-user limits. Codex Business and Enterprise are credit-based with their own spend controls. Both are real accounting machinery. The question is whether you need it.

How Two Lanes Became One

My setup went through three architectures in a week, and each one deleted a mechanism.

Version one was a wrapper. A PATH shim with a directory-scoped launcher: inside one project tree it verified a company OAuth token, checked the token was attested to the right domain, injected it into the environment and refused to launch at all if any of that failed. Outside that tree it did the mirror-image check for the personal account. Session-start hooks in five settings files warned if the wrong identity was in play. It was fail-closed, carefully built, and it worked.

It was also several hundred lines of shell whose entire job was guessing which credentials a directory deserved. Two token stores, two verifiers, two guards, four login scripts, and a launch path that had to be ordered correctly or every unattended worker exited 127. I built it. I removed it.

Version two was a machine boundary. Personal accounts on my laptop, company accounts on the cloud dev box. Same separation, zero code — the boundary was physical, and physical boundaries do not need a verifier.

Version three came from the company itself. Use the reimbursed personal subscriptions on the cloud machine too, so all work runs through subscriptions the engineer chose, and consider retiring the centrally-managed plan entirely — because there is no real benefit to centralized accounting for it.

End state: one account per vendor, everywhere. Laptop and cloud box. No shim, no guard, no branch, nothing to verify. The most engineering I did in that whole sequence produced the least value, and deleting it made everything more correct.

There is a real argument on the other side, and I want to state it fairly. Centralized seats give you provisioning, offboarding and audit — when someone leaves, one admin action revokes their access. Reimbursed personal subscriptions give you none of that; you get a receipt and a trust relationship. If you have compliance obligations or meaningful headcount churn, that machinery earns its keep. For a small team of senior people, it is overhead with a login flow attached.

The quieter argument for reimbursement is the one that actually won: it puts the choice of harness with the person doing the work. A budget lets an engineer pick the tool. A seat picks it for them.

What I'd Do On Monday

Concretely, if you run agents for a living and someone else is paying:

Find your real ceiling. Open the usage panel on both vendors before you plan anything. On Codex you get published numbers; on Claude you get your own consumption history, which is the same information the hard way.

Assume headless and interactive share it. Budget your unattended runs against the same weekly cap you use by hand, because that is literally what happens.

Set the overage cap to remaining headroom. Not total budget. The meter starts at zero and ignores what you already bought.

Turn auto-reload off if the card is not yours. Buy one tranche, take the volume discount, let the hard stop be your circuit breaker.

Ask which seat, not whether there's a plan. A Standard seat and a Max 20x differ by a factor of sixteen and nobody volunteers that.

Isolate unattended spend if agents starve your sessions. A console API key is a separate wallet at the same rates. Same money, different bucket, and your own terminal stops competing with your loops.

None of this is exotic. It is just that the pricing pages describe subscriptions and the docs describe APIs, and almost nobody writes down the part where an unattended agent quietly drinks from the cup you are holding.

The review agent that died at 7:21 came back at 10:15. It had been running against a plan I thought had room, doing work I thought was billed somewhere else. Both of those beliefs were free to hold and expensive to keep.