The £99 AI Spend Autopsy, explained

7 July 2026 · 5 min read

The AI Spend Autopsy is a one-off report. You connect read-only keys, we analyse 30 days of usage across your providers, and you get an itemised list of savings with the arithmetic shown. It costs £99. If we find less than £99 a month in savings, you get the £99 back.

This post walks through a real autopsy — run on Halcyon Labs, the fictional company that ships as TokenLedger's demo dataset — so you can see exactly what the findings look like and how each number is derived. Halcyon spends about £4,473 a month across OpenAI and Anthropic. The autopsy found £1,114.78 a month, roughly 25% of spend.

| Finding | Confidence | Monthly saving | | --- | --- | --- | | Stop paying for retries | HIGH | £213.81 | | Right-size frontier-model workloads | MED | £488.20 | | Cache the repeated prompt prefix | MED | £269.95 | | Batch the nightly workload | MED | £142.82 | | Total | | £1,114.78 |

Finding 1: Stop paying for retries — £213.81/mo, HIGH

For every key, TokenLedger computes an hourly spend baseline from the trailing month. On 6 July, Halcyon's evals-nightly key spiked to 23.6x its hourly baseline. Nothing shipped that day and no dataset grew. That shape — one key, one night, an order of magnitude above baseline — is a retry loop: a failing request being resent, each attempt billed in full, until a timeout gives up.

The formula: saving = spike excess × 0.8. The spike excess is what the key spent above its baseline during the event — £267.26. We apply a 0.8 recovery factor because some elevated spend during an incident is legitimate (genuine reprocessing, partial successes). That gives £213.81.

Why is this monthly? Because the loop is in a nightly job and the bug does not fix itself. Until someone adds backoff and a retry cap, every transient failure re-runs the same billing event. Fixing it once removes the class of spend, not one instance of it.

This is HIGH confidence: the spike is directly observed in the billing data, and the only assumption is the recovery factor.

Finding 2: Right-size frontier-model workloads — £488.20/mo, MED

One of Halcyon's keys sends gpt-4.1 a high volume of requests with long inputs and very short outputs — the signature of classification, not generation. Labelling a support ticket does not need a frontier model. gpt-4.1-mini handles short-output classification at roughly one fifth of the price.

The formula: saving = spend × price delta × 0.4. The identified classification traffic is £1,525.63 a month. The price delta between gpt-4.1 and gpt-4.1-mini is 0.8 — the mini model costs 80% less per token. We then apply a 0.4 adoption factor: we assume only 40% of the traffic actually migrates, because some requests will turn out to need the larger model and no migration ever reaches 100%. £1,525.63 × 0.8 × 0.4 = £488.20.

MED confidence: the traffic shape is observed, but "this workload will perform acceptably on the smaller model" is an assumption you should verify with an eval before switching.

Finding 3: Cache the repeated prompt prefix — £269.95/mo, MED

Two of Halcyon's keys send the same multi-thousand-token system prompt on every single request — and the usage data shows zero cached tokens. Prompt caching is opt-in on Anthropic, and nobody opted in. Cached input tokens are billed at a 90% discount, so an identical prefix resent thousands of times a day is the cheapest fix on this list: it is a one-line change.

The formula: saving = input cost × cache discount × 0.6. The input-token cost attributable to the repeated prefix is £499.91 a month. The cache discount is 0.9. The 0.6 adoption factor covers cache misses: entries expire between requests, prefixes occasionally change, and not every request window keeps the cache warm. £499.91 × 0.9 × 0.6 = £269.95.

MED confidence: the repetition and the zero cache reads are observed; the realised hit rate is an estimate.

Finding 4: Batch the nightly workload — £142.82/mo, MED

Halcyon runs jobs at 02:00 every night — embedding refreshes and evals — through the standard synchronous API, at full interactive rates. Nobody is awake. Nobody needs the answer in 800 milliseconds. Both major providers offer a batch API at a 50% discount for workloads that can wait up to 24 hours, and a job scheduled for 02:00 is the definition of a workload that can wait.

The formula: saving = night spend × 0.5 × 0.5. The recurring night-window spend is £571.28 a month. The first 0.5 is the batch discount. The second 0.5 is the eligibility assumption: only half the nightly volume is assumed to migrate, since some jobs feed morning dashboards with tighter deadlines. £571.28 × 0.5 × 0.5 = £142.82.

How confidence works

Every finding carries HIGH, MED or LOW confidence, and the labels have rules.

  • HIGH: the waste is directly observed in billing data, with at most one conservative assumption. Counts toward the headline number.
  • MED: the pattern is observed, but the saving depends on a stated adoption or performance assumption. Counts toward the headline number, with every assumption printed next to the figure.
  • LOW: plausible, but resting on something we cannot see from usage data alone. Listed separately as "possible, verify first" and excluded from the headline. If a LOW finding pans out, it is upside we did not charge you for.

Halcyon's £1,114.78 is one HIGH finding plus three MED findings. No LOW findings are in that number.

The guarantee

The autopsy costs £99, once. If the HIGH and MED findings total less than £99 a month in savings, tell us within 30 days and we refund the £99 in full. On the demo dataset the findings came to 11x the fee; on real accounts with more than one engineer and a retry loop, the ratio is usually not close.

Connect read-only keys, wait a few minutes, read the arithmetic. That is the product.