Sample report. Fernhill Analytics is a made up name, but the usage data comes from a real client, shared with their permission and with figures lightly altered.
Estimated savings on your AI bill
That's 18% of the $109,100 you spend on AI each month. If every fix lands at the high end, savings reach about $42,900 a month (39%). Over a year, that's roughly $238,000 to $515,000.
Some fixes act on the same spend, so we count that overlap once. That's why this total is a little less than the sum of the findings below.
Fernhill already does a lot right. The product assistant caches its shared context, some bulk work already runs through the Batch API, and nothing runs on a retired model, though one older model is due for an upgrade.
Most of what's left to save is in three places:
Every source of AI spend you shared with us, from July to September 2026:
More than two thirds of your model spend goes to Opus, your most expensive model. That's where most of the savings are.
API and Claude Code usage by model($98,100, monthly average for July to September)
ModelPer month Share
Versions within each model family are combined, since new versions came out during the quarter. Sonnet 4.5 is shown on its own because it's the older model we suggest moving off. Shares are rounded, so they may not add to 100%.
All AI spend by product($109,100, monthly average for July to September)
ProductPer month Share
Seats aren't tied to a model, so the model chart covers API and Claude Code usage only. Shares are rounded, so they may not add to 100%.
Ranked by the low end of each range. Ranges are per month. Overlapping savings are counted once in the headline. Open any finding to see the detail behind it.
Where it shows up
What to do
Run jobs that don't need an instant answer (nightly summaries, reports, evals, bulk tagging) through the Message Batches API.
Why we are confident
Method and assumptions
75% to 100% of the background work you named can move, at 50% less per token.
A two week sprint would cache the support triage prompt, then build a quality check from real tickets so routine ones can move to Sonnet without worse answers. Nightly reports would move to the Batch API, and the doc-parser workspace would move off Sonnet 4.5 to Sonnet 5.5 after a side by side check. Developer plans and seat cleanup need no engineering time.
In your report, names appear as anonymous codes that match a file that stays on your machine.
Every figure is computed by code from your usage data and checked twice: once against an independent calculation, and once against Anthropic's own monthly totals. Savings use your actual billed prices, including any discounts you already receive.
Prompt caching, model routing, trimming long prompts and moving off Sonnet 4.5 act on the same spend, so the headline counts their overlap once: each saves a share of what is left after the others.
We only read usage, cost and seat totals. We never saw your prompts, outputs, or payment details. You ran our usage export tool on your own machine with your own keys, reviewed what it found, and sent us the file. Your key was never shared with us. You can revoke it as soon as the export is done. Raw data is deleted 30 days after this report. Anonymous totals, never your name, may be used to compare you with similar companies, only if you agree.