Adam Ferree — Building with AI. At scale.

A little perspective.

478,000

copies of The Great Gatsby

Approximate token volume, using Anthropic’s published 72,000-token measurement.

View source

11.5×

GPT-3’s Wikipedia corpus

Compared with the historical 3-billion-token English text dataset in the 2020 GPT-3 paper.

View source

196 years

of nonstop reading

A rough extrapolation of five hours per 100,000 tokens. Twenty-four hours a day.

View source

The numbers behind the numbers

These comparisons put recorded token volume into perspective. Tokens measure pieces of information processed by a model; counts differ by model and content. The comparisons do not measure unique writing, completed projects, human work replaced, or equivalent computing cost.

The recorded usage

This is a dated snapshot of recovered Codex and Claude Code logs, generated 2026-09-29 19:15 UTC. It covers the current server, recovered retired-server archives, and laptop. Home PC history, deleted logs, and ordinary ChatGPT and Claude web chats are not included. It is the recovered record, rather than a complete lifetime account total.

Recorded token categories
Category	Tokens
Uncached input	496,376,968
Cached input read	33,191,219,946
Cache writes, 5-minute	193,575,185
Cache writes, 1-hour	409,712,570
Output, including reported reasoning	129,300,751
Total	34,420,185,420

96.4% of the total consists of cache reads: previously supplied context reused across requests. The output count includes reported reasoning and tool-related output; it is not all published prose.

Copied sessions and repeated streaming records are deduplicated. Codex cumulative counters are converted to request increments, inherited fork balances are excluded, and cached input is separated from uncached input. Claude usage records are deduplicated by message and request identifiers.

The dollar estimate

$29,833.84 is the estimated standard API equivalent at the September 29, 2026 price snapshot, applying recorded input, cache, and output categories and applicable long-context pricing. 177,486,651 tokens have no matched price and contribute to token totals but not the dollar estimate.

The separate $206,776.01 comparison hypothetically prices all input without cache discounts. It is not the headline estimate, an invoice, or money saved. Neither estimate includes subscription fees, tools, images, taxes, or fast/priority surcharges. The underlying logs do not establish the billing mode of every request, so these figures are not a subscription billing audit.

Price references: OpenAI API pricing and Claude API pricing. Historical usage is repriced at the dated snapshot; historical bills are not reconstructed.

What everyday AI tasks look like

Published examples and their scope
Example	Token count	Basis and source
Read one PDF page	1,500–3,000 text tokens, plus image tokens	Typical page estimate from Anthropic. Responses and transcription output add tokens.
Generate a square image	272 / 1,056 / 4,160 image output tokens	Low / medium / high quality at 1024×1024 for GPT Image models before GPT Image 2. Prompt input is extra. OpenAI.
Read The Great Gatsby	72,000 tokens	Anthropic’s 2023 demonstration loaded the entire novel into Claude.
Process an hour of audio	115,200 input tokens	Calculated as 3,600 seconds × 32 tokens/second from Gemini’s documentation.
Process a two-hour movie	About 1.89 million input tokens	Calculated as 7,200 seconds × 263 tokens/second for Gemini’s static video processing. Google.
Read a substantial codebase	A 1-million-token context	Google illustrated Gemini 1.5’s capacity with over 30,000 lines of code. This is an input-capacity example, not a website-building budget. Google, 2024.
GPT-3’s Wikipedia corpus	3 billion tokens	English text dataset in Table 2.2 of the 2020 paper. Not today’s complete Wikipedia or the cost of rebuilding it.

Comparison formulas: total ÷ 72,000 for Gatsby; total ÷ 3,000,000,000 for Wikipedia; total ÷ 100,000 × 5 ÷ 24 ÷ 365.25 for years of reading. Displayed figures are rounded. The reading speed is a rough estimate from Anthropic, not a reading-time measurement of these logs. Text, image, audio, and video tokens have different pricing and are not interchangeable units of work.

Download usage snapshot

What I do

Two jobs. Deciding what to build, and then building it.

Most AI engagements stall between the strategy deck and a working system. I do both halves, so nothing gets handed off and lost.

Voice agents that answer a real phone number.

Production telephony, not a demo widget. Inbound and outbound, call recording and transcripts, escalation paths, and the carrier and compliance layer underneath. One of mine currently answers a public help line around the clock, in a regulated category, without a human in the loop.

Retell. Telnyx. Twilio. SIP.

iOS apps, through review.

SwiftUI, StoreKit 2 in app purchase, subscriptions, and the App Review process end to end. Ten live on the App Store today.

The infrastructure under it.

Nginx, TLS, DNS automation, Stripe billing, monitoring, nightly backups. 23 domains and 31 certificates, administered solo.

Consulting: what not to build.

Half of what gets proposed as an AI project is a workflow problem wearing a costume. I will tell you which half yours is, in plain language, before anyone writes code.

Proof

Do not take my word for any of this.

Four claims on this page are independently checkable in under two minutes. Here is how to check them.

Download it RevenuePack on the App Store Native iOS, live subscription billing, approved through Apple review. One of ten apps published under my name. Open the listing Buy it revenuepack.com Live commerce, not a waitlist page. Stripe subscription billing and StoreKit purchases, both running against production accounts I set up and reconcile. Open the checkout Use it protectindiana.org Free scam checking for Indiana consumers, backed by a 435,000 domain blocklist. Paying business customers fund it. Consumers never pay. Try the tool Watch it amfindiana.com The portfolio page reports live telemetry off the production server, refreshed about once a minute. If something breaks, you will see it before I fix it. See the servers

Apps live on the App Store. Designed, built, submitted, and supported by one person.

Results

I am also the operator who has to live with the numbers.

My day job is asset protection operations for a 332 associate Walmart supercenter, leading a team of fifteen. Every supervisor on that team was promoted out of the hourly ranks.

That is the part most AI consultants are missing. I do not hand a system to an operations team and leave. I have been the operations team. I know what gets ignored at 6am on a short staffed Saturday, and I build for that shift instead of for the demo.

So the tools I build at work get measured the way everything else in a store gets measured. On the P&L, against plan, every period.

0.86 from 4.99

Facility injury rate, one fiscal year. Chain average that year: 5.2.

$87K from $502K

Accident cost, same year.

1 from 9

Workers compensation claims.

1.55%

Shrink held against a 1.54% plan, through a full remodel year.

These are employer numbers, so they are not published anywhere I control. They are a reference call away, and I will give you the reference.

Background

Twelve years in loss prevention. Then I taught myself to ship.

Internal and external investigations as an hourly associate. Supply chain asset protection in a 24/7 powered equipment facility. Asset protection assistant store manager for a high volume metro store, including leading that store through the 2020 Indianapolis civil unrest. Six years running a licensed general contracting company with full P&L ownership, from first estimate to final invoice.

None of that is a software resume, and that is the point. I came to AI engineering from the side of the business that has to absorb whatever technology gets dropped on it.

Everything I build now, I also run. The billing, the servers, the App Review rejections, the 2am pager. There is no team behind me to escalate to, which means nothing ships unless it actually works.

Tell me what you are trying to automate.

If it is a bad idea, I will say so on the first call. That is usually the most valuable hour of the engagement.

adam@adamferree.com