By the numbers

• Seeded on day one with 80 days of prior chat history, then 68+ days of live interaction since the first build (April 1).
• Experience can be bootstrapped or built cold — the primary operator seeded from history; three other users started from zero, all seeing value within days.
• 4 active users across distinct domains, plus several more intermittent.
• 5,267 turns served since launch — over 6,400 counting the seeded history.
• 600+ active memories, 1,500+ archived — a corpus that compounds, not a cache that resets.
• 500+ commits, managing a real book daily.
• Field notes drawn from 130+ verified behaviors — we published the ones that anonymize cleanly, with receipts. Far more where that came from.
• ~$0.42 per turn, all-in, across 5,267 turns since launch — ~$2,200 of total API spend, every layer of the system included, not just the answer. The builder's own dense, system-building sessions run higher (~$0.48); a typical user runs ~$0.34.

---

A note on cost

That ~$0.42/turn is the total system cost — not just generating the answer, but the machinery that compounds experience: routing, nightly consolidation, governance audit, memory synthesis. A plain chatbot's cheap turn is inference only; nothing accrues between sessions, so there's nothing to pay for. We're not comparing the same product.

That $0.42 is the lifetime average across every turn since launch. The recent rate is higher — about $0.65/turn blended — and we'll be plain about why: the system now runs on the top-tier model (a recent upgrade that raised both quality and cost), and the builder's own dense, system-building sessions run ~$0.90/turn and pull the blended average up. A typical user runs ~$0.50, and the heaviest real-domain user ~$0.38. That recent blended number is the honest one to plan against — and it is the most expensive these economics will ever be.

It's retail API pricing paid by a solo builder, on a system whose cost-optimization work is only just beginning. Three levers all point down: committed-use pricing (a real product gets discounts off list — a lever not yet pulled), engineering (cost optimization is the work starting now — the first lever shipped this week), and amortization (a user's nightly passes cost the same whether they take 5 turns or 50, so per-turn cost falls as their usage deepens; and the genuinely shared infrastructure spreads across the user base as it grows — currently 4 active plus several intermittent).

Which makes per-turn the wrong unit twice over. Take "Back to Amsterdam." On a flat-memory model, re-establishing that context costs a long re-explanation every session — many turns, every time, forever. That cost doesn't disappear; it moves onto the human and recurs. Here the system carried the thread and resumed in one turn.

So the real comparison isn't per-turn vs per-turn. It's one all-in turn that picks up where you left off versus ten cheap turns rebuilding what was already said — plus the user's time and patience, which no metric captures.

We pay more per turn to spend fewer turns, run the whole learning system inside that number, and move the re-priming cost off the human entirely — at the worst unit economics this system will ever have. And we're driving it down on every lever at once.