Claude Agent Cost Audit · $1,500 flat

Your Claude agents now have
a separate bill.

On June 15, 2026, Anthropic moved programmatic Claude usage — the Agent SDK, claude -p, Claude Code GitHub Actions, third-party agents — off your subscription's limits and onto a separate monthly credit, metered at full API rates, with no rollover. If you run Claude agents in production, the question stops being “does it work” and becomes “what does each run cost, and what halts when the credit runs out.” This audit answers both, with numbers from your actual code.

Book the cost audit general code audit → or email hello@sprayberrylabs.com

Flat fee, no surprises · Baselined from your real metered usage · NDA-friendly

What changed June 15

The facts, so we're working from the same page:

Who this hits

Agents on cron or CI

Scheduled jobs, GitHub Actions, pipelines calling claude -p — the quiet automations nobody has priced per-run.

Products on the Agent SDK

If your product's features run on Agent SDK calls, your unit economics change on the 15th — whether you've modeled it or not.

Agencies running client automations

Each client workload now has its own meter. Margins that worked under a flat subscription may not survive API rates.

Anyone who got the claim email

Anthropic is emailing subscribers to claim their credit. If you're not sure what it means for your stack, that's the audit's first question.

What the audit covers

A focused pass over the code that runs your agents, with the June 15 economics applied to it:

Why it's different

01

We run agent fleets ourselves

This studio runs on its own AI workforce — agents on schedules, budget gates, cost tracking, the lot. The audit applies operating experience, not a framework skim.

02

Every finding verified

Each finding is checked against your real code — file, line, and symbol confirmed to exist — by a deterministic gate before it's allowed into the report. No hallucinated problems.

03

A senior signs off

Twenty years of engineering reviews, prioritizes, and stands behind every finding. You're not getting raw model output — you're getting a senior engineer's audit, accelerated.

What you get

A finding looks like this

High jobs/daily-digest.yml · agents/digest.ts:41

Daily digest re-sends the full document corpus on every run

What: The scheduled digest agent rebuilds its prompt from scratch each morning — ~180k input tokens of unchanged reference documents — with no prompt caching and a frontier-tier model doing summarization a mid-tier model handles identically.

Impact: At metered API rates this one cron job consumes roughly half of a Max 5x monthly credit by itself. Under the subscription it was invisible; after June 15 it starves every other automation by mid-month.

Fix: Cache the static corpus with prompt caching, send only the day's delta, and drop the summarization step one model tier. Same output, ~90% lower per-run cost; add a per-job budget gate so a regression like this pages someone instead of draining the pool.

Illustrative example — shown to give you the format and depth, not a real client's finding.

Want to see the audit format end-to-end? Two real reports: one that found & fixed a live command-injection → and one that found & fixed a ReDoS →. Every finding mechanically verified against the source.

Price & terms

$1,500 flat — delivered in days 50% to start, 50% on delivery

One fixed fee agreed up front — no hourly drift, no scope games. With the change now live, an audit produces an exact picture from your actual metered usage — no more estimating, the numbers are on your statement. Your code stays yours, an NDA is welcome before the first call, and nothing you share is ever used to train any AI model. If the audit points to work worth doing, it folds cleanly into a Sprint — but there's no obligation, and no upsell baked into the findings.

Common questions

What changed on June 15, 2026?

Anthropic moved programmatic Claude usage — the Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party agents — off your subscription's usage limits onto a separate monthly credit (about $20 on Pro, $100 on Max 5x, $200 on Max 20x), metered at standard API rates with no rollover. When the credit is exhausted, automated requests stop — unless overflow billing is enabled, in which case they bill at API rates.

What do you need from me to start?

Read access to the code that runs your agents — a GitHub or GitLab invite, or a zip — and recent usage numbers if you have them. That's enough to begin.

Can you still audit agents now that the change is live?

Yes — and now the data is real. With the billing change live, an audit produces an exact cost picture from your actual metered usage: per-run cost at the new rates, model-tier fit, caching and prompt waste. No more estimating — the numbers are on your statement.

What if my setup is already fine?

Then you get a short report saying so, with the numbers to prove it — documented confidence instead of a background worry. The audit is a diagnosis, not a rewrite pitch.

Is my code kept confidential?

Yes. I'm happy to sign your NDA before the first call, your code is never used to train any AI model, and access is removed once the audit is delivered.

Book the cost audit

Know your number before the meter starts.

Send the repo that runs your agents and a sentence on what they do. You'll have the per-run economics, the risks, and the fixes — in days.

No spam, ever. Thomas reads every one personally and follows up — usually within a day. See our privacy policy.