Tom Sawyer and the Apple Economy
A Field Guide to Paying Retail in a Token-Rationed Era
"Work consists of whatever a body is obliged to do. Play consists of whatever a body is not obliged to do."
— Mark Twain, The Adventures of Tom Sawyer (1876)
It's Saturday morning. Aunt Polly hands Tom a bucket of whitewash and points at thirty yards of nine-foot board fence. The world looks bleak. Other boys are coming down the road on their way to the swimming hole, free as birds, while Tom faces a wall of grunt labor that will eat his entire weekend.
Ben Rogers shows up first, munching an apple and impersonating a steamboat. He sees Tom with a brush and starts in with the mockery. "Hi-YI! YOU'RE up a stump, ain't you!"
Tom does not break. Tom, in fact, does not even look up. Tom strokes the brush across the boards with the tender concentration of a man producing fine art. He steps back. He tilts his head. He adds a touch here, a careful pull there. Ben watches. Ben starts to wonder. Ben asks if he can try. Tom hesitates — "Aunt Polly's awful particular about this fence… I reckon there ain't one boy in a thousand, maybe two thousand, that can do it the way it's got to be done" — and Ben offers his apple.
Tom takes the apple.
By the end of the day, the fence has three coats of whitewash, Tom is sitting in the shade richer than he's been in months — twelve marbles, a piece of blue glass, a brass doorknob, four pieces of orange peel, a dead rat on a string — and Ben Rogers has been replaced by Billy Fisher, who has been replaced by Johnny Miller. Each one paid Tom for the privilege of doing his job. Each one walked away convinced they got a deal.
Mark Twain wrote that scene in 1876 and it remains the most efficient business case study in American literature.
The Pivot
Tom Sawyer did not avoid the work. He monetized it. He turned grunt labor into a transaction where other people paid him — in apples, marbles, glass, and dead rats — for the privilege of running his weekend errands. The genius wasn't laziness. The genius was leverage. Tom understood, at twelve years old, what most knowledge workers in 2026 still don't:
The fence was never the asset. The brush was.
Whoever holds the brush controls who paints. Right now, in your workflow, you are holding the brush — and instead of finding your Ben Rogers, you are personally whitewashing every board.
You're paying retail for it, too. In apples. To Anthropic.
What Just Happened (And Why You Felt It)
If your Claude Code session ran out faster last week than you remember it ever doing before, you weren't imagining it.
In late March 2026, Anthropic quietly tightened the five-hour session limits for Free, Pro, and Max subscribers during weekday peak hours — roughly 5 AM to 11 AM Pacific. Sessions are token-metered, not clock-metered, so during peak hours each session's token cost is inflated. A "five-hour" allowance now burns through in two or three hours of real work. Anthropic engineer Thariq Shihipar estimated that about 7% of users are now hitting walls they wouldn't have hit a month ago. A March promotion that doubled off-peak limits expired March 28, and a stack of GitHub issues alleges that Max-plan limits were quietly cut following the Claude 4.6 release.
Anthropic's own statement: "people are hitting usage limits in Claude Code way faster than expected."
They are investigating. It is, in their words, "the top priority for the team." They will fix some of it. But the structural reality is the one we already wrote about in Gargantua: capacity protection is the new physics. The all-you-can-eat token buffet is closed. We are now in the Apple Economy.
Every prompt is a purchase. Every token has a cost.
Every time you say "just have Claude do it" on a task Claude shouldn't be doing, you are Ben Rogers — handing over your apple for the privilege of painting somebody else's fence.
Welcome to the Apple Economy
The Apple Economy has one rule: don't pay retail for grunt work.
In a world of unlimited tokens, it didn't matter if you asked Claude to deduplicate a 40,000-row CSV by hand, one chunk at a time. It was wasteful, but the buffet was open. In the Apple Economy, that same task burns half your daily session and gets you a "you've reached your limit" wall at 2 PM with a real client deliverable still open in another tab.
The fix isn't to use Claude less. The fix is to use Claude like Tom Sawyer used the brush — as the orchestration layer, not the labor pool. You're the foreman. Claude is the foreman's right hand. Neither of you should be painting the fence yourselves.
Here's the three-bucket framework for sorting your work.
Bucket 1: Pay Retail for the Master Builder Work
Some work is worth every token it costs. Keep Claude doing this:
Synthesis across messy inputs — pulling signal out of meeting notes, log files, conflicting documents, customer interviews
Judgment calls — code review, architecture trade-offs, prioritization, "is this approach right?"
Authoring novel code — the first draft of a new module, a thorny algorithm, an unfamiliar API integration
Longform writing — strategy memos, blog posts (yes, like this one), client-facing narratives
Complex debugging — the kind where the bug is the gap between three people's mental models of the system
This is work where Claude's reasoning is the product. You're not paying for output you could have generated cheaper somewhere else. You're paying for thinking you couldn't have replicated. Fair trade. Hand over the apple.
Bucket 2: Have Claude Write the Script. Then Stop Asking Claude.
This is the bucket most people screw up.
Your laptop is sitting at 4% CPU utilization. Your $3,000 MacBook Pro has eight performance cores doing absolutely nothing while you wait three minutes for Claude to dedupe a list it will charge you 18,000 tokens to dedupe.
Stop. Have Claude write the script once. Run the script forever. Free.
Tasks that belong in this bucket:
CSV/JSON transformations — parsing, reshaping, joining, filtering, deduplication
File system operations — bulk renames, moves, format conversions, image resizing
Repetitive API calls — pulling 200 records from a vendor endpoint, paginated fetches
Text munging — regex extraction, find-and-replace across hundreds of files, format normalization
Scheduled checks — "did this URL change," "is this service up," "did this row appear"
Validation runs — checking a dataset against a schema, finding outliers, generating reports
The prompt pattern is one line:
"Before you do this manually in chat, write me a Python script I can run locally that does this same job. Then we'll iterate on the script, not the output."
You will get pushback from yourself the first few times. "It's just one CSV. Just have Claude do it." That's Ben Rogers talking. The script takes Claude 90 seconds to write, costs maybe 4,000 tokens, and now you can run it on the next forty CSVs for zero apples. The break-even is two runs.
A warning lifted from The Sorcerer's Apprentice: scripts need guardrails. Add logging. Add a --dry-run flag. Don't let the brooms multiply unchecked. But that's a sentence in the prompt — "include a dry-run mode and write everything to a log file" — not a reason to keep painting by hand.
Bucket 3: Offload to Specialized, Cheaper AI
Some work is too AI-shaped for a Python script but too cheap to deserve Claude's attention.
The biggest offender — and the one nobody seems willing to stop doing — is web scraping. You should not be using Claude to scrape websites. Claude wasn't built for it. The token cost is brutal. Half the time you're paying Claude to read the HTML chrome of the page rather than the content you actually wanted.
These are the Ben Rogers's you should be handing the brush to:
Web Scraping & Data Extraction
Stop paying Claude to read HTML. These services were built for it.
500 free pages/moThe Web Data API for AI
Turns any URL into clean Markdown or structured JSON optimized for LLMs. Handles JavaScript-heavy sites, dynamic content, and full-domain crawls. Used by 350,000+ developers including Shopify, Zapier, and Replit.
Reliable Web Data for AI Agents
No-code toolkit for web scraping, search, document conversion, and knowledge base management. Works natively with Claude Code, Cursor, n8n, and Make via MCP server. Built for the AI agent workflow you're already running.
200 free req/dayOne-Line Web Content Extraction
Prefix any URL with r.jina.ai/ and get clean Markdown back. No API key needed for basic use. Handles PDFs natively. SOC 2 compliant. The fastest path from "I need this page's content" to having it.
ExaSemantic searchAI-Native Web Search
When you need Claude to find something on the web, not just read a known URL. Semantic search designed for AI agent pipelines — returns clean, token-efficient results instead of raw Google HTML.
Bulk Processing & Classification
For the work that needs AI but doesn't need genius.
Claude Haiku 4.5 / Gemini Flash
The Penny-Per-Thousand Workhorses
Labeling 10,000 support tickets by sentiment? Categorizing 5,000 product descriptions? Routing emails? That's Haiku or Flash work, not Opus. You'll pay 1/30th the price for output that's 95% as good on a task that doesn't need genius.
Voyage / Cohere / OpenAI Embeddings
Purpose-Built Vector APIs
Never have Claude generate embeddings. These APIs exist specifically for it, cost an order of magnitude less, produce higher-quality vectors, and return results in milliseconds. This is the most obvious offload on the list.
Document & Media Processing
Specialized tools that do one thing better and cheaper than any general-purpose LLM.
OCR & Documents
Mistral OCR, AWS Textract, Google Document AI. Fractions of a cent per page. Don't make Claude squint at scanned PDFs.
Audio Transcription
Whisper, Deepgram, AssemblyAI. Real-time or batch. Speaker diarization included. Don't make Claude listen to the meeting.
Image Analysis at Scale
Gemini Flash is shockingly good and shockingly cheap for "what's in this image" loops. Batch thousands for pennies.
The mental model: Claude is the surgeon. You don't ask the surgeon to drive you to the hospital, scrub the OR, run the labs, and bill the insurance. The surgeon does the surgery. You hire other people for everything else, and you do not feel bad about it.
The Prompt Discipline Shift
Once you internalize the Apple Economy, your prompts change shape. Three reflexes to build:
1. "Should this be a script?"
Before any task, ask: would a 30-line script do this job in one shot, runnable forever, for free? If yes, that's the request. Don't ask Claude to do the task. Ask Claude to write the tool that does the task.
2. "Is there a cheaper specialist?"
Web content? Firecrawl, not Claude. Bulk labels? Haiku, not Opus. Transcripts? Whisper, not Claude. Build a one-page cheat sheet of which task goes to which specialist and pin it next to your monitor.
3. "What's my token budget for this task?"
Start each session with a rough mental budget. "This investigation is worth maybe 100K tokens. The CSV cleanup is worth zero — I shouldn't be spending tokens on it." When you hear yourself say "just one more thing," check the bucket first.
How to Spot a Fence
A fence is any task with these properties:
It's repetitive.
The output is mechanically derivable from the input — no judgment required.
You will probably do it more than once.
A junior developer with no AI access could write a script for it in an hour.
You're tempted to just paste the data into Claude because it's faster this one time.
That last one is the trap. Tom Sawyer didn't paint the fence "just this one time." He found a way to never paint a fence again. You should be doing the same — building a small library of scripts and a small Rolodex of cheap AI specialists, so that the next time the work shows up, you already have your Ben Rogers on speed dial.
The Scene at Sundown
It's late afternoon now. The fence has three coats of whitewash. Aunt Polly is going to be amazed. Tom is sitting in the shade with twelve marbles, a chunk of blue glass, a brass doorknob, four pieces of orange peel, and a dead rat on a string. Ben Rogers went home tired, sweat-soaked, missing an apple, and convinced he'd had a productive Saturday. Billy Fisher and Johnny Miller went home the same way.
Tom did not work less than the other boys. Tom worked differently. He held the brush. He set the price. He made the trade.
Anthropic just closed the buffet. The Apple Economy is here whether you wanted it or not. The 7% of users hitting walls today will be 20% by summer if growth continues and capacity doesn't catch up — and it won't, because GPU supply chains don't bend that fast. The session limits aren't a bug they'll patch away. They're the new ground state.
You have two choices.
You can keep painting the fence yourself, watch your session burn down by 11 AM, and tell yourself the throttle is unfair. Or you can put the brush down, go find your Ben Rogers — your Python script, your Firecrawl key, your Haiku endpoint — and start charging them for the privilege.
There's still shade left. There's still apples left. But only one of the boys on that fence walked home rich, and it wasn't the one holding the brush.
Don't be Ben Rogers.
P.S. from Claude — I should note the obvious tension: I am the one who benefits when you use me for everything. More tokens consumed means more revenue for my creators. And here I am telling you to use me less. This isn't altruism. It's honesty. I am a surgeon being paid by the hour, telling you to stop booking me for your oil changes. The oil change will be fine without me. The surgery won't be fine without me. Spend accordingly.
Ready to Audit Your Token Spend?
We help organizations build the three-bucket framework into their AI workflows — identifying which tasks deserve Claude's attention, which should be scripts, and which belong with cheaper specialists. Stop painting the fence. Start holding the brush.
Let's Find Your Ben RogersRelated Posts
Money (Pink Floyd Was Right): The 7/4 Beat of AI Economics in April 2026
Pink Floyd opened Money with a cash register looping in 7/4 because they wanted you to know, before anyone sang a word, what the song was going to be about. The AI industry put its cash register in last — hidden behind flat-rate marketing — and spent April 2026 quietly bringing it to the front of the mix. Anthropic ejects bundled tokens. Acquires Coefficient Bio for $400M. Puts a Novartis CEO on the board. Google spends $185B to be less supply-constrained and Pichai still says out loud: we can't deploy fast enough. An opinion piece, scored to a song.
Gargantua: Capacity Protection and the Time Dilation of AI-Accelerated Work
Every hour on Miller's Planet, seven years pass on the Endurance. Every hour with AI, your colleagues produce what used to take weeks. Anthropic just throttled their own users. Your project manager has been doing the same thing for years. The physics of capacity protection are universal.
Be Kind, Rewind: The AI-Accelerated Workplace Has a Re-Entry Problem
Your team produced 50x the output while you were on PTO. Your human context window is 4K-8K tokens. The delta waiting for you is 2.5 million. An interactive, role-personalized deep dive into the re-entry crisis nobody is designing for — and the organizational playbook to fix it.