The subsidy was the moat.

We have published two pieces here about how little human time it takes to ship a product. Eleven hours for one, about a day for another. Both were measured rather than estimated, and both left out the compute.

Over the last thirty days, my model usage came to $15,008.02 at list API rates.

It sat on a subscription. That gap is the only thing keeping anyone inside a particular vendor's tooling — and the labs are in the process of giving it up.

What they are actually selling

Every one of these companies runs two ledgers. A seat is an OAuth credential scoped to that vendor's own agent and its own backend. An API key is console-issued, points at a public host, and bills per token. A ChatGPT Business subscription does not include API usage; a SuperGrok subscription carries no xAI credits; a Claude subscription is not an Anthropic API key. What a third-party harness gets from the OAuth flow is not a key either.

Anthropic has drawn that boundary hardest, and their documentation is unusually direct about it. OAuth authentication exists "to support ordinary use of Claude Code and other native Anthropic applications." The binary "must not be modified." Developers building anything on top are told to use API keys instead. And since 4 April 2026, subscription quota no longer covers third-party harness traffic at all: the same account can still make the call, but it bills as extra usage rather than drawing on the plan you paid for.

So the rule is exactly what it looks like. Your plan's economics exist inside their harness and nowhere else. That is not an accidental side effect of an anti-resale policy. It is the policy, written down, and it is the clearest published statement of what these subscriptions are actually for.

Which is fine, until you try to buy it properly. An Anthropic Enterprise seat includes Claude, Claude Code and Cowork, and then: "Usage isn't included in the seat fee. Every token your team uses … is billed at standard API rates on top of your seat cost." That seat reportedly lists at $20. It replaced a flat plan that ran up to $200 per user. A bundle became a login.

The gap those seats were hiding is large. SemiAnalysis ran coding on the plans until the weekly caps and repriced the tokens at list: up to roughly $8,000 a month of API-equivalent value on a $200 Claude Max seat, $14,000 on a $200 ChatGPT Pro seat. Quesma put a headline on it from the buyer's side — same tokens, same model, up to 40× the price.

And that is the mistake

Here is the part I have not seen anyone say plainly. The flat plan is the only lock-in these companies have.

It is not the model. Model leadership rotates on a timescale of months, and everyone reading this has watched it change hands. It is not the harness either. A coding agent is a few thousand lines of orchestration — a loop, a tool protocol, a context strategy, a diff applier. Real engineering, some of it very good, but not a semiconductor fab. Whoever ships fastest leads it.

What holds a serious user in place is that the same work costs an order of magnitude less inside the vendor's own client. It is the only reason anyone tolerates a harness they would not otherwise choose.

Move that customer onto the meter and it is gone. On list-price tokens the official agent and every other agent cost the same to run, and the official one has to win on merit alone, permanently. Nobody wins that forever. The moment a better harness exists — and one will, because the substrate moves too fast for any client to stay ahead of it — no bill stands between the user and the switch.

And lock-in is only the first thing to go. Loyalty goes with it. On a flat plan a developer defends their vendor, argues for it in meetings, learns its quirks and stays through a bad month, because leaving costs them real money. On the meter none of that is true. There is nothing to be loyal to: every provider bills the same way, at rates that fit in a spreadsheet, and the comparison takes an afternoon.

Brand goes next. "We're an Anthropic shop" is a meaningful sentence when it describes an economic position. When the invoice would be identical anywhere else, it describes a habit. Habits survive one disappointing model release. They do not survive two.

What is left is a feature race on the harness itself — and a harness is the most copyable artifact in this entire stack. There is no patent on a good context strategy. Everyone reads everyone else's changelog, the good ideas propagate within a quarter, and the lead belongs to whoever shipped most recently. That is not a moat. It is a treadmill, and they have just stepped onto it voluntarily, having removed the one thing that meant they did not have to run.

So the rational architecture is easy to state: isolate your day-to-day work from the platform underneath it. Own the loop. Treat models as interchangeable parts with different prices and different competences. That layer already exists in several open implementations, and most of them are behind the first-party agents today. Behind is temporary. The labs have removed the only durable reason to wait for them, and arranged for their customers to fund the catching up.

Compliance revenue is real — a bank will pay per head forever for a vendor to hold accountable. But that is a compliance business, not an ecosystem business, and it is no defence against a platform team that puts three providers behind an internal gateway.

What the meter reveals

There is a second reason the meter is dangerous for them, and I only saw it in my own numbers. Of that $15,008, 83.4% ran on Opus and 16.2% on Fable — the second-most and the most expensive models on the rate card. Effectively the whole month ran on frontier reasoning.

That is not a defensible mix. It is what a harness produces when nothing in the loop is paying attention to cost, and there are three reasons for it.

1. The model is fixed when the session starts. A long agentic run is architecture decisions sitting next to file reads, grep sweeps and log triage. You choose once, at the top, and the cheap steps inherit the expensive choice.
2. Investigation runs on text search instead of an index. Watch one of these agents understand a codebase and it is grep, find and shell pipelines, reading files whole to answer questions a symbol index answers in a token or two.
3. Context is re-read rather than retrieved. The 18.3 billion cached tokens against 38 million billed ones tell you how much of this is the same material moving past the model again. Caching makes that cheap. It does not make it necessary.

At least half the frontier share was work a mid-tier model would have done indistinguishably, and a slice below that belongs on a cheap tier or on no model at all. I have not run the experiment, so I will not put a figure on it. The direction is not in doubt.

One caveat: this exceeds the ceiling SemiAnalysis measured for a single seat, and I am not claiming to have beaten their method — different window, different mix, their stress-tested ceiling against my month across many projects. The two agree on the order of magnitude, and that is the argument.

The uncomfortable inference

Put those together. If serving this work cost the vendor anything close to what it bills for it, the harness would not look like this. You do not leave a 5× cost lever untouched in your own client for eighteen months while paying the difference yourself.

The counter-argument is fair: these teams optimise for capability, latency and trust rather than their own cost of goods, and a harness that silently downgrades your model is a support nightmare. So the inference is directional, not proof.

The consequence holds either way. On a flat plan that inefficiency is the vendor's problem, and they have chosen not to solve it. On the meter it becomes the customer's largest controllable line item — and the customer will solve it, by routing each step to the model it actually requires. Which means routing across vendors. Which means leaving.

The labs have handed their most sophisticated users the motive and the budget to build the thing that makes those labs interchangeable.

The opposite bet

I should declare an interest, because this is written from inside the problem. Vela is a harness, in the sense this piece has been using the word. It sits above the models, decides what runs where, and does the part the model on its own cannot.

It has been model-agnostic since the first version. Every step in the pipeline is configured with its own model; the main deployment runs on Cloudflare, with configurations that can be deployed against Google, Anthropic and OpenAI. Anything exposing an OpenAI-compatible endpoint is reachable.

The reason was not the argument in this article. It was sovereignty. A client who needs inference inside their own tenancy — their own Azure OpenAI deployment — has to be able to have it, because the rest of the product already made that guarantee, and a model layer that broke it would have made the guarantee worthless. Portability was the price of a promise we had already made. That it also happens to be the thing this whole article is about is luck, of a kind.

Underneath it was a deliberate bet: that the model becomes a commodity and the value moves to what sits above it. I think the labs know this too. The difference is that they cannot charge for their harness and we can — because ours encodes what to extract, what to verify, what to escalate to a human and what to refuse. A model on top of a model. That does not come in the box with any of them.

None of which describes my own development loop, whose numbers open this piece and which has no such discipline whatsoever. Different system, different constraints — but I would rather publish the gap than pretend it is not there.

Three things worth doing on Monday

Find out which ledger your pilot ran on. If the engineer who ran it was signed into a personal plan, the cost you were shown is not the cost you will pay, and the multiple is not small.

Measure your model mix before you negotiate. If your usage looks like mine did, you are paying frontier rates to run text searches — worth more than the discount you were going to ask for, and it changes what you should be negotiating.

Buy the harness you can leave. Ask what it would take to move your engineering workflow to a different model provider, and require the answer in days. If nobody knows, you do not have an AI strategy. You have a supplier, on a rate card they revise unilaterally.

And if your engineers keep a frontier subscription on a personal card because the sanctioned path costs ten to forty times as much for identical work, that is not a discipline problem. It is your pricing model reproducing itself as shadow IT.

Where this ends

Nobody rips out a frontier vendor in a quarter. What happens is quieter: classification, retrieval, first-pass patches and test generation move to whatever is cheapest and adequate, one step at a time, and the expensive model is kept for the turn that can fail in public. That is most of the tokens leaving without anyone announcing a migration.

The flat plan was the one thing that made staying cheaper than leaving. It is being withdrawn from exactly the customers with the scale to notice, in favour of a rate card that prices the vendor's own tooling identically to everyone else's. What remains is a bet that their harness stays the best one.

No harness stays the best one. That is what a harness is.