Baton: an experiment in spending AI wisely

There’s a bill almost nobody is looking at yet. Today, using AI to write code is paid by the month, at a fixed price — but that price is propped up by a subsidy that is unlikely to last. When the bill shifts to pay-per-use, as it already has in other markets, running every task on the most powerful (and most expensive) model will stop being affordable for an independent developer.

The obvious reaction is a sensible one: not everything needs the most expensive model. A commit, a summary, a routine task — a more modest model does those well; save the big muscle for design and the hard decisions. Baton is the experiment we’re building to explore exactly that: routing each job to the right model, using the discipline StrayMark already produces as a compass. It’s in development, it’s experimental, and we’re assembling it in the open. Its story runs in three chapters — and none of them ended where we expected.

The experiment, in three acts

Coherence first — before saving, check the bearing. The trap showed up immediately. If you make who does the work cheaper but the work has already drifted from the plan, you saved nothing: you made the mess cheaper and faster to produce. So the first thing Baton built wasn’t a money-saver but a coherence check: does what’s being built still match what was decided? Put to the test on a real project, it caught a genuine drift that human review had let through — without touching a single model, without spending more.

The honesty of the number — what the dry run would have spent. When it finally looked at the money, a tempting headline appeared: a huge saving, of more than ninety percent. And there Baton did the opposite of showing off — it told on itself. More than half of that saving rested on guesses — the system guessing what kind of work it had in front of it when it had nothing to go on. The number that mattered wasn’t the saving; it was the doubt written next to it. A bright figure on top of a pile of guesses is still a pile of guesses.

Ask the one who already knows — what the author already knew. The fix turned out to be almost free. Instead of guessing what kind of work each task is, why not ask the person who creates it? Whoever opens the work already knows whether it’s design, implementation, review, or a mechanical chore — they know it for free, in the moment, for the cost of typing one word. The cheapest and most reliable signal in a system is usually the one someone already knows and was never asked to write down.

Why this is StrayMark

Baton started as a question about money and ended up being, once again, the same old bet. It isn’t about using more AI, but about the AI you already use spending wisely — and about having a tool honest enough to say “I don’t actually know” instead of faking a confident number. It’s the same conviction we set out in Staying oriented in strange days: that the human stays in command, not on the sidelines.

The question was never how much the machine can save. It was whether the saving rests on things someone is sure of — or on guesses with a good face.

An experiment in the open

Baton doesn’t actually route anything yet: for now it recommends, measures, and learns. It may graduate into StrayMark’s core, or it may not; that’s the nature of an experiment run in the open. We share it because the journey teaches more than the destination: coherence before saving, honesty before the headline, and asking before guessing.

If you want the technical detail we deliberately left out here, the three chapters are on StrayMark’s development blog: coherence, the honest number, and asking the author.