The AI Token Economy
AI can give you and the people you lead super powers, but it comes with hidden managing responsibilities you didn't sign up for.
Motivation
Thought experiment: Imagine you're assigned a freelancer with these weird characteristics.
On paper, this collaborator arrives with encyclopedic knowledge, works instantly, and accepts tasks without hesitation. But they carry three peculiar flaws.
strength (?)
Fast & well-read
Knows about everything you ask them, and can produce results fast without breaks.
strength (?)
Accepts any task
No task too big or too niche — it will take a swing at whatever you hand it.
strength (?)
Tireless
Never stops, never sleeps, unless you stop paying.
flaw
Never asks questions
Takes tasks literally, fills gaps with assumptions, never pushes back.
flaw
Bills by the word
No flat day rate. Every prompt, tool call, and correction adds to the tab.
flaw
It remembers nothing.
Every new conversation starts from zero. Yesterday’s work never happened.
Lead One, Upskill Many
It's an LLM — but if it were a human, how would you manage it?
Directing an AI agent requires the same delegation and review rigor as onboarding an external specialist.
You can opt out of managing it. The people who don’t will simply out-produce you.
door 1 · the individual
Human with AI
5–20× individual throughput by directing synthetic labor.
door 2 · the organization
Leader of humans with AI
Upskilling whole teams into fluent directors of automated systems.
The Four Eras of AI
New technology gives rise to new disciplines for humans to cope.
From Prompt Engineering in 2023 to Context Engineering in 2024, Harness Engineering in 2025, and Token Economics today.
2022–23
chat workflow
Prompt Engineering
Phrasing, few-shot examples, and asking clearly.
2024–25
agentic workflow
Context Engineering
AGENTS.md, skills, and compact structured memory.
2025–26
autonomous workflow
Harness Engineering
Tools, deterministic tests, and guardrail verification.
2026 →
emerging
Loop Engineering?
Self-improving autonomous feedback cycles.
Foundation
Token Economy — the currency of every era
The Invisible Invoice
For a real freelancer you need a business case. AI just bills you.
A €3,200 consultant invoice requires purchase orders and approvals. AI accumulates identical costs silently across thousands of micro-prompts.
Invoice #0042approved
Slide design€1,800
Alignment€900
Revisions€500
Total€3,200
AI · No Invoicenever quoted
+€0.80+€1.20+€0.60+€2.10+€0.90+ thousands more…
Cumulative Total€3,200
Outcome Per Euro
You're now accountable for a budget you never signed up to manage.
Defaulting to flagship frontier models for routine chores burns capital. The economic winner optimizes outcome per euro.
Premium model
5× the priceQuality delivered99%
Outcome per euro
financial winner
Cheaper model
1× the priceQuality delivered95%
Outcome per euro≈5×
93%
At 93%, both models clear the bar — the cheaper one wins ≈5× outcome per euro.
2–10×
Today’s prices are subsidized — vendors sell intelligence below cost to build adoption. When subsidies expire, outcome per euro decides who survives.
What Is a Token
Tokens are the basic unit of data for LLMs — every one is billed.
LLMs calculate statistical probabilities over numerical tokens. Every word, character, and prompt translates to billable units.
Insanityisdoingthesamethingoverandoveragainandexpectingdifferentresults.
≈ 24 tokens (approximate — every model splits differently)re-sent across a 50-turn agent session ≈ $0.0060 at Opus 4.8 input rates
Every token is billed 4 ways
- inputevery word you send — and re-send on every turn
- outputevery word generated in return and reasoning traces
- cache writestoring recurring context blocks for reuse
- cache readreusing cached blocks — ~10% of base input rate
Live Context Breakdown
Most tools hide the meter — this is what seeing it looks like.
A typical 200k session breakdown: system prompts, conversation history, tool payloads, and retrieved documentation.
claude-code — /context
$ /context
Sonnet 5
claude-sonnet-5
79.5k/200k tokens (40%)
Estimated usage by category
- ⛁System prompt9.4k · 4.7%
- ⛁System tools42.9k · 21.5%
- ⛁MCP tools3.9k · 2.0%
- ⛁Custom agents703 · 0.4%
- ⛁Memory files407 · 0.2%
- ⛁Skills2.2k · 1.1%
- ⛁Messages20.0k · 10.0%
- ⛶Free space87.5k · 43.7%
- ⛝Autocompact buffer33k · 16.5%
CLI vs Web Tool Obscurity
The meter is hidden on purpose.
Obscurity is a product pattern: frictionless initiation, habitual workflows, and surprising monthly invoices.
01 The initiation
Unlimited on the house. The intelligence feels like magic and the bill is invisible — so you build the habit.
02 The hook
Now you are fluent — and the meter is gone. Out of credits? Just top up. The real rate never shows.
Managing Context
Context window is the model's attention span.
Tokens accumulate in the context window. As the window fills, attention degrades and critical instructions slip away.
Plain chat25% · fresh
Boot-up
Agentic coding85% · full
Boot-up
9 AM · fresh12 PM3 PM5 PM · running on fumes
The Lost Middle
How key instructions get lost in the middle.
Models attend sharply to the start and end of context, but the middle blurs into an attention valley.
System prompt (Primacy)“You are an expert coder…”
[msg 12: read utils.js]
[msg 13: modify api.js]
[msg 14: read config.json]
Latest request (Recency)“Why is the build failing?”
Serial-position effect: Models recall the first and last exchanges sharply, while the middle blurs into an attention valley.
The Fullness Gauge & Levers
Past ~half full, the model starts losing the plot.
Below 50% is the smart zone. Above 50%, accuracy degrades. Above 75%, hallucination risk spikes. Three levers: Monitor, Compact, Hand off.
~50% · hallucination onset
Smart zone · groundedDanger zone · hallucinates
85%
Danger zone · Running on fumes
past ~50% what “losing the plot” looks like
- · misremembers the middle
- · contradicts earlier steps
- · invents files & APIs that don’t exist
~illustrative threshold, not an exact number — but degradation well before the limit is measured across frontier models
Monitor · /context
Watch the gauge before it degrades — catch it while it is still grounded.
Compact · handoff
Summarize where you are into a clean handoff, then start fresh from it.
Cost of Agentic AI
One small bug fix compounds to ~73,000 tokens of context.
Agentic loops re-transmit the entire conversation history on every single tool call and reflection step.
System prompt· rules, tools, memory
+35,00035,000
User query· the bug report
+1,00036,000
Thought #1 & Tool call· read file #1
+4,20040,200
File #1 contents· entire 400-line file
+12,00052,200
Thought #2 & Tool call· read helper module
+4,20056,400
File #2 contents· second full file
+9,00065,400
Thought #3 & Patch apply· write and verify patch
+8,05073,450
Total context hauler~73,450 tokens
The Compounding Cost Curve
The same bug fix, priced across models.
You are billed for tokens processed on every turn, not window capacity. A routine bug fix produces a 14× price swing across models.
DeepSeek V4 Pro$0.11
Claude Sonnet 5$0.62
Claude Opus 4.8$1.54
~14× spread for identical output. At 50 fixes a day, that is roughly $6 vs $77 per developer per day — before anyone has shipped a single feature. This is why model routing matters.
50
Model Tiers
My real model stack: one workflow, many models.
Route each phase to the most economical capable model: flat-rate chats for ideation, frontier reasoning for architecture, fast models for code execution.
Phase 1
Ideation & Refinement
Web chat · flat-rate
- · Simple research → Gemini Flash
- · Brainstorms → Gemini Flash / Pro
- · Voice notes → DeepSeek V4 Pro
Phase 2
Specs & Planning
Claude Code CLI · Desktop
- · PRD & Architecture → Opus 4.8
- · Deep research → Sonnet 5 → Opus
- · Type checks → Sonnet 5
Phase 3
Execution
All CLI
- · Commits → Gemini Flash
- · Implementation → Sonnet 5
- · Exploration → Haiku 4.5
Tools & Tips
Three levers to improve your token economy
Combine automated tooling (rtk, compression, linting) with structural context steering and disciplined team habits (edit don't append, session hygiene).
Automated
- Headroom: Compresses agent transcripts before they hit the LLM.
- rtk: Compresses CLI command outputs with low overhead.
- Model Routers: Route dynamically to the most economical capable tier.
- caveman: Enforces concise output saving up to 75% on generation.
Agent Steering
- Memory Files: Persist durable rules via AGENTS.md, CLAUDE.md, and profiles.
- Specialized Skills: Separate operational workflow guides from base instructions.
Team Habits
- Edit, Don’t Append: Fix prompt mistakes at the root rather than stacking corrective turns.
- Session Hygiene: Clear finished context rather than paying stale cache re-reads.