Follow AI spending
to the work behind it

01 · FIND THE WORK BEHIND THE TOTAL

Start with who, when, and what ran

The monthly view shows totals and changes over time. The ticket list makes the spending concrete: which task ran, who started it, which model it used, and its estimated cost. Separate rates for new input, cache writes, cached input, and output explain how the tokens become dollars.

Here, Maya’s APP-58 specification run stands out. Follow that run before deciding whether the answer is a different model, a different method, or justified investigation.

02 · INSPECT THE INVESTIGATION

Three full reads deserve a closer look

APP-58 used 9.52M tokens across 64 investigation calls ($37.97), 2.406M across 18 specification calls ($11.01), and 1.02M across 8 acceptance-review calls ($5.26).

Inside the investigation, api-schema.json was read in full three times. A strong prompt can repeat this expensive discovery when the repository does not expose a useful, smaller contract surface.

The trace gives you a cause to investigate, not just a larger total.

04 · MATCH THE MODEL TO THE WORK

Reserve premium models
for exceptional complexity

A model such as Fable 5 or GPT-6 Astra should earn its place through the difficulty of the task. Routine investigation and well-specified changes deserve a lower-cost route that meets the same engineering requirements.

For APP-58, GenieWorks suggests reviewing the model choice: a bounded investigation of existing export paths and schema ran entirely on Fable 5. Try Sonnet 4.6 for equivalent work and compare the specification, evidence, and acceptance coverage.

Repricing the recorded token categories gives $37.97 in potential API savings per run. This isolates model rates; it does not include a saving from narrower file reads.

Standard Anthropic API rates, checked September 21, 2026. Same token counts and cache mix; actual tokenization, calls, cache hits, and repair work can change. No cheaper run has been measured in this example.

05 · TEST THE IMPROVEMENTS

Change one thing
Check the outcome

Test the model suggestion with the original task, context, and workflow. Compare the resulting specification and acceptance coverage, including any retries or repair work needed to reach the same standard.

Separately, test selecting the relevant schema objects before reading the API contract. Keep the model fixed for that comparison; a line limit alone would not bound a single-line JSON file.

The model saving is a rate-only estimate. Savings from narrower reads remain unmeasured; do not add the two together.