understudy

Local Claude spend audit

Lower your Anthropic bill.

The local skill that analyzes your product's integration and figures out ways to get the same job done better, cheaper, and faster.
Open source. Runs locally. Yours forever.

curl -fsSL https://raw.githubusercontent.com/UnderstudyLabs/understudy-agent-tools/main/install.sh| bash -s -- --lower-my-ant-bill
book time with our team
Find Anthropic callsOptional bill evidenceRe-count Opus tokensAudit cache missesRank cheaper routesEvaluate open weight models
local audit
installer --lower-my-ant-billalias --lower-my-anthropic-billskill lower-anthropic-billscan app/api, workers, evalsbills optional email/browser scopetokens Opus tokenizer drift flaggedcache 8 static prefixes foundnext approve spend before provider calls

Why this matters

Claude should stay where it wins.

Repeated summarization, extraction, classification, routing, and repair should earn the premium route. The audit builds the ledger first, then asks before spending.

18% / 25%

of Sonnet cost on sales slices

5.2x

lower latency on bounded ops

4.4x / 50x

cheaper than Sonnet and Opus labeling

Common failures

Same bill. Three usual causes.

The install starts local and turns the vague bill into a route ledger: what is overpowered, what is uncached, and what can be owned.

A PhD in the mailroom.

A premium reasoning model is sorting envelopes: extraction, labels, summaries, repair, and routing.

Try Sonnet with tighter prompts first. If quality still has headroom, run GEPA against a measured eval and keep Opus only where it wins.

Busted cache, busted bill.

A timestamp, request id, shuffled tool list, or short prefix can break cache hits before you notice.

Move stable tools, system text, schemas, examples, and documents before volatile fields. Confirm with cache usage, not hope.

Don't rent, own.

Repeated work becomes an annuity for the frontier provider when the route never graduates.

Try an open-weight model on Fireworks through Understudy. If it clears the eval, ramp it behind a fallback.

Optional evidence

Find the bill, then find the route.

Repo inspection is the default. Billing email, invoices, exports, and browser usage pages are optional read scopes when you want real spend hotspots.

Email receipts

If you approve it, the agent can search only Anthropic billing messages and pull dates, totals, and visible usage fields.

Billing console

If the coding harness supports browser inspection, you can open the usage page and let it record model, token, and cache hotspots.

Exports first

CSV, JSON, screenshots, or pasted totals stay local and give the cleanest ledger when browser or email access is not available.

First run

Scan. Count. Rank.

Map Anthropic SDK calls, fetches, model aliases, retries, and owners.

Re-count expensive Opus routes and check prompt-cache breakpoints.

Propose cache, batch, Haiku, OpenAI, and local candidates with approval gates.

Talk it through

Book time with our team.

Bring one expensive route, rough monthly volume, current model names, and the first workflow you would rather own.

open in Cal.com

Share savings

Anonymous receipts. Leaderboard coming soon.

When an eval-backed claim proves lower spend, the agent can send a metrics-only savings report back to Understudy. No prompts, traces, repo names, company names, or contact details.

/api/lower-my-anthropic-bill/savings
Sources checkedOpus tokenizerClaude cachingOpenAI caching

Disclaimer

Independent informational site.

This independent site is for informational purposes only. Understudy Labs is not Anthropic, is not affiliated with Anthropic, and is not endorsed, sponsored, or backed by Anthropic. Anthropic, Claude, and related names are trademarks of their respective owners. Any pricing, token, caching, or savings discussion is an estimate based on public information and local analysis, not financial, legal, procurement, or technical advice. Verify current provider documentation, pricing, contracts, and usage terms before making changes to production systems or model routing.