Six models a month and no tested exit.

The question

Is the ground I am standing on still ground? The Economic question is whether the conditions that make this profitable today will make it profitable tomorrow, when the prices are subsidized, the models retire on a schedule nobody publishes, and the harness changes under you between one quarter and the next. On this build the question is which model to use this week. On the long clock it is whether a practice built on a provider's terms survives that provider changing them.

Beside it sits the Ethical line the model reads under the same family: how concentrated is my practice in any one provider's tooling, and is there a working fallback if access were cut. Not a fallback I believe in. One I have run.

What the line says

Four readings. Distinct models that carried my work in a month: one in October 2025, five to seven a month since March 2026, six this month, each arriving with a new name and a retirement date for the last. Share of tokens by provider: one provider carried all of it a year ago and about a fifth of September's; the other carried nothing then and carries the rest now. No fallback test has been recorded, so the number has no tested exit behind it.

Share of exchanges by harness: one tool carried everything for nine months, since July the other has carried most of it, and this month the work is split. And of all the input I send, the share the provider served from cache: about half for nine months, nine-tenths now. The last three describe one migration from three sides: whose model, whose tool, and how much context the tool carries on every turn.

None of this is spend. Subscription harnesses report what they used, not what it cost. And a new model number is not a new capability. It is churn, counted.

The three questions

Every research post runs the same three rows. Which of the nine it names, how good the evidence is, and whether the thing is being measured.

What this cannot say:

1. What any of it cost. Volume is not spend. The pricing rows that would turn one into the other exist and most of them are unreviewed.
2. Whether the migration was a choice or a drift. A better harness, a rate limit, and a habit all produce the same curve.
3. Whether nine-tenths cached is cheaper or just heavier. A cached token is lighter per token and a sign the tool is carrying more context on every turn. Both are true at once.
4. Whether I could actually leave. Concentration is a number. An exit is a rehearsal, and none has been run.

What I am doing about it

The status says what is missing. These are the moves, on my own work, that would supply it. Each is either read automatically off the footprint or asked of me on a cadence, and each costs something to keep doing.

1. Auto · QuarterlyRun one real workload on the alternate provider every ninety days, and record the flag. The locked signal carries a tested-fallback flag. It is unrecorded because I have not done the drill, not because the drill is hard.Costs: a quarter-day on the tool I like less, and a reading that may say the exit is a hope.
2. AutoAttach a price to the volume. Provider pricing tables, versioned, so volume becomes spend normalized against output, which is the Economic signal the model actually asks for.Costs: pricing rows that go stale monthly and must be reviewed before they are trusted.
3. AutoWrite the substitution test for each binding before I need it. The kit's category map asks, for every tool in the shop, what a replacement must pass. Recorded in the settlement, it is what makes a migration a decision instead of a weekend.Costs: writing the test before I need it, which is the only time I will.
4. Asked · QuarterlyWhat roles or hours am I no longer paying humans for that I would have a year ago? The Ethical family's labour question. The provider share above is where those dollars went.Costs: an answer I would rather keep to myself, written down.

What I keep