The free 5-minute audit.
Your AI bill, audited in five minutes.
One command, run locally against your own usage export. Metadata only, never message bodies. You get a banded spend X-ray and honest pointers to what's recoverable — never a savings figure until it's proven on your real traffic.
npx recovea scan --sample # sample data, zero setup
npx recovea scan usage.csv # your own exportHow it works
Three steps. No account, no API key, nothing uploaded.
One command, on your machine
Export usage from your provider's console (OpenAI: Usage → Export, Anthropic: Console → Usage) and point the scan at the file. No API key, no account, no install that outlives npx.
We never read message bodies
The scan reads token counts, model ids, request counts, and dates. Prompts, completions, and keys aren't in the export and are never asked for. Local-first means even that metadata stays on your machine. Nothing reaches us.
A banded spend X-ray, and where to look
Your 30-day spend comes back as a band (never a point figure), net of quality, to be verified. We label that band modeled: the export is measured, the month ahead is projected from it. Plus non-dollar pointers to where the money concentrates. Recovea claims $0 in savings until it's proven on your real traffic.
What comes back
A unified spend number, a banded 30-day projection, where spend concentrates, and honest pointers to what's recoverable — no savings figure until it's proven. Every figure is re-derivable by hand.
Recovea spend X-ray · priced against rpl-2026-08-05 (re-derivable)
$764.00 total · 178.0M in / 86.0M out · 136,000 requests
~ $3,820.00/mo (band $3,247.00–$4,393.00) · from a 6-day export, net of quality, to be verified
Model Requests Cost Share
gpt-4o 42,000 $500.00 65%
claude-opus-4-8 5,000 $190.00 25%
gpt-4o-mini 80,000 $39.00 5%
claude-haiku-4-5 9,000 $35.00 5%
Where to look (pointers, not promises. Recovea claims $0 savings until proven on your traffic)
- Output-heavy: 72% of spend is output tokens (priced up to 5x input). Trimming verbose/looping generations targets this; measure to confirm.
- Premium-model spend: 90% runs on top-band models (gpt-4o -> gpt-4o-mini, claude-opus-4-8 -> claude-haiku-4-5). Routing the simpler calls to a smaller sibling could apply, only if quality holds (needs an eval, not a guess).
- No cache reuse detected: 100% runs on cache-capable models but the export shows no cache-read tokens. Prompt-cache / exact-cache hygiene targets repeated prefixes; measure on real traffic to size it.
⚠ 1 model(s) not in rpl-2026-08-05, excluded from the total above:
text-embedding-3-small (50.0M in / 0 out)
Every figure = your tokens times the published rpl-2026-08-05 rates. Check it yourself:
cost = input/1e6*in_rate + output/1e6*out_rate + cached/1e6*cache_rate
Want this measured continuously on your real traffic instead of a one-off export?
Point your tools at the gateway: npx recovea install · recovea.ai
What this costs and whether it is worth it at your spend: recovea.ai/scanAn illustrative sample of the report format, generated from the built-in sample export (--sample), not output from real data. Your numbers are computed from your own export, on your own machine, and the report can come back "not worth it."
What the numbers are, and what they are not.
- A band, never a point figure. Your 30-day spend prints as a range. A precise forward number would pretend we've metered traffic we haven't; the band absorbs window-length and price-snapshot drift.
- $0 in savings, until proven. The scan never prints a recoverable dollar figure. What's recoverable comes back as pointers — where money concentrates and which lever could apply — because Recovea claims $0 in savings until it's proven on your real traffic.
- "Modeled," never "verified." The band projects a month from a shorter export, so we label it modeled: the export itself is measured, the month ahead is not. Verified means proven on your real traffic behind a per-route quality gate, and the scan has metered a file, not your traffic. "To be verified" means exactly that.
- Net of quality. The projection is labeled net of quality: any lever that would trade output quality for cost stays off the table until an eval proves it holds. Savings that cost you quality aren't savings.
- Re-derivable by hand. Every figure is tokens × published rate against a frozen price list (rpl-2026-08-05). Run with
--explainto see the arithmetic behind each line.
If this isn't worth it for you, we'll say so.
And we say it here, on this page, where you can hold us to it — not inside the tool. The scan reports your numbers and points back here. It used to decide for you from a threshold compiled into the package, which drifted from this page the moment either changed: it told anyone under $15K a month not to pay us, while the entry plan is $49 a month. One place for that answer means one edit to correct it. An audit you can't fail is a sales pitch.