The £1,940 month, and the wiring that closes it
Eleven days of receipts, then the architecture that brought the next month back under control.
We logged the receipts from 1 to 11 May, named the failure modes, and shipped the wiring fix. The case study is the working not just the answer.
Receipts first, then the architecture, then the honest limits, then the five lessons.
The setting
The setting
In the first eleven days of May 2026, The Glucose Never Lies Ltd burned £1,940 on Anthropic API. At that rate we were on a path to roughly £64,000 a year. We caught it, fixed it overnight, and the projected forward saving is around £270,000 over five years.
We are a small team. John Pemberton is the founder and operator. Phillip Hayes is the technical director, part-time. Anjanee Kohli holds the creative line.
We ship GNL Grace, a clinical educational advisor for people with Type 1 diabetes (educational tool, not a medical device), plus the wider GNL site and consultancy work. Most of our build runs on Claude Code, which is a working developer surface, not a query API. Through April and into early May, our daily spend climbed.
The receipts
The receipts
The first eleven days of May, day by day: 1 May, £75. 2 May, £540. 3 May, £382. 4 May, £288. 5 May, £220. 6 May to 11 May, around £465 across the rest of the run-up to mid-month. Approximately £1,940 in eleven days. Annualised at that rate, that is on a path to £64,000 a year.
Educational tool attestation. GNL Grace is an educational tool, not a medical device. Clinical decisions sit with the person and their diabetes care team.
The wiring fix
The mystery and the wiring
Months earlier, Phillip had taught John to think like a developer. Not “learn to code” in the bootcamp sense; something subtler. How to lay out a project so that a tool can find what it needs without burning a fortune on retrieval. That groundwork made the next move possible. Phillip’s server architecture upgrade in early May genuinely cut Grace’s per-query cost by 85% on the v2 deploy, 6 May 2026. The same architectural pattern, ported to Claude Code itself, was the wiring fix that closed the £1,940 month.
Projection
What it would look like for you (projection, not a quote)
| Frame | Figure | What it means |
|---|---|---|
| Own-spend at trial | £1,940 | Eleven days of GNL Anthropic spend on Claude Code, before wiring fix. |
| Own-spend annualised | ~£64,000 | If the same pattern had continued for a full twelve months. |
| Annual saving thereafter (projected at GNL session volume) | £20,000 to £60,000 | Forward projection from the £1,760 lever the wiring fix closed. |
| Five-year forward protection (projected at GNL session volume) | ~£270,000 | Compounding effect of the wiring fix at GNL session volume. |
Every forward-saving row is projected at GNL session volume; the figure for a client engagement is scoped per engagement, not anchored to a public number. The working sits in docs/TOKEN_OPTIMISATION_PLAYBOOK.md.
Five lessons
The five lessons
- Build, then wire, are two separate jobs.
- The cheapest thing is usually already there.
- Cost attribution beats cost reduction.
- Token-optimisation is mostly a session-design problem.
- Free local tooling does the heavy lifting.
Skin in the game
Skin in the game
We did this with our own money on our own product. £1,940 of trial-and-error sits on John’s personal Visa for May 2026, reimbursed via the director’s loan account. Annualised at that rate, that is on a path to £64,000 a year, or £270,000 over five years. We caught it in week eleven. We fixed it in three hours.
This is what we mean when we say Via Negativa is a consultancy with skin in the game. We did not learn this on your dime. We learned it on ours. We have already absorbed the cost, walked the path, debugged the dead ends.
Honest limits
What this case study does NOT claim
We do not claim every team will see the same lever. The wiring fix worked here because the substrate was right (a heavy read-class workload, a content surface big enough to index, a session pattern that rewarded retrieval discipline). We do not claim the forward saving is guaranteed. We do not claim Claude Code is the only path; we use it because it is the working developer surface for our stack. We will tell you on the first call if the substrate is wrong for the lever.
Talk to us
Talk to us
Via Negativa is the consultancy arm of The Glucose Never Lies Ltd. We work with small AI-product companies whose Claude Code bill is climbing faster than they expected. No public price sheet; scope first, figure second.
Reference this case study by name: “GNL £1,940 month”.