AI Economics
What tokens, agent runs, subscriptions, and model choices actually cost, measured on real workloads rather than list prices.
Each chapter can be read on its own. Within a chapter, the posts are in reading order.
Token Cost
Prompt caching, effective cost per million tokens, and how to attribute spend to a team or tenant.
-
Why AI Coding Tools Are Getting Cheaper: Prompt Caching Explained
Updated:Prompt caching discounts repeated input, but the savings depend on exact prefix reuse, model-specific retention, and enough reads to repay each cache write.
-
Prompt Caching Is a Harness Design Problem, Not Just an API Toggle
A 186-request test shows why cache savings depend on stable tools, deterministic prefixes, deliberate breakpoints, routing keys, and measurement inside the agent harness.
-
So What Is the Effective Cost for a Million Tokens on Frontier Models?
The sticker price on a model card is not what you pay. Measuring cache-hit ratios from months of real usage across Sonnet 5, Opus 5, and GPT-5.6 puts the effective cost per million tokens 70-95% below list price.
-
Every Way to Attribute Cost on Amazon Bedrock, Live-Tested
A working map of every mechanism Amazon Bedrock has for attributing inference cost — resource tags, identity tags, and request metadata — with config examples, and where SaaS billing has to leave the AWS billing layer entirely.
Agent Tooling
A cost-justified order for agent tools, where the harness wins back margin, and what a hybrid agent costs once it is optimized.
-
API → Connector → Browser → Computer Use → Human: A Cost-Justified Hierarchy for Agent Tooling
Computer use is 45× more expensive than a structured API for the same task. Eight harness configurations across browser, desktop, and hybrid showed costs from $0.03 to $5.44 on the same task — the hierarchy, the numbers, and why implementation beats model choice.
-
I Had a Working Agent at $3.41 a Run. Here's Where the Money Was Actually Going.
Same task, different harness configurations. Caching the inner loop saved 29%. Convergence rules saved 54%. Eliminating the LLM from the browser loop entirely dropped cost from $3.41 to $0.03. The harness is where the margin lives — and the harness goes all the way down.
-
The Test Isn't Whether the Agent Picks the Right Tool. It's What Happens When the Tool Fails.
I built a hybrid agent expecting the routing to be the demo. The actual demo turned out to be what happened when a tool call failed and the outer model retried with a different approach — twice in two runs.
-
The Cross-App Workflow Cost $5.44. Here's How We Got It to $0.17.
A hybrid agent running browser research + desktop spreadsheet creation cost $5.44 baseline. Optimizing each leg independently — Playwright for the browser, CLI-first prompt for the computer use desktop leg — brought it to $0.17. The browser leg did most of the damage.
Subscriptions
Why AI subscriptions behave like usage plans, how reset windows shape use, and why autonomous agents break flat-rate pricing.
-
AI Subscriptions Are Secretly Usage Models
Claude, ChatGPT Codex, Cursor, Devin, Copilot, Gemini, Perplexity, and Grok all look like subscriptions. The real product is the reset policy, the overflow rule, and how much autonomous work you can squeeze out before the meter shows up.
-
Reset Windows Are Product Design
Five-hour resets, rolling 24-hour restores, daily caps, weekly quotas, and monthly credits are not billing details. They are the actual behavior design of AI subscriptions.
-
Autonomous Agents Break Flat-Rate Pricing
The problem with AI subscriptions is not that vendors are greedy. It is that autonomous agents turn flat-rate pricing into a bad fit for the actual cost shape of the work.
-
How To Optimize Agent Subscriptions Without Getting Tricked
The best way to get more out of Claude, Cursor, Codex, Devin, Copilot, Gemini, and Perplexity is not clever prompting. It is operational discipline around scope, context, retries, and reset timing.
Model Market
Matching models to tasks, the intelligence-versus-price frontier, open-weight catch-up, and where routing decisions live.
-
The Right LLM for Each Step in an Agent Pipeline
Updated:Match model tier to cognitive load. A fast, single-call path with pre-fetched context cuts a 25-30 second two-call response to under 10 seconds.
-
The Pareto Frontier Is the Model Market Map
Model leaderboards tell you who is smartest. The Pareto frontier tells you which models are still rational choices once price enters the room.
-
Benchmark saturation is real. The price collapse is an illusion.
Industry claims models are plateauing and prices are crashing. One is true but misframed. The other is a measurement trap.
-
Open Weights Catch the Last Frontier, Not the Moving One
Across 19 model snapshots, open-weight leaders reached an earlier proprietary frontier in 24–58 days on coding, intelligence, and agentic benchmarks—but remained behind the live frontier.
-
The Gateway Controls the Request. The Router Chooses the Model.
Model routing is growing because model choice has become a runtime decision. Gateways govern traffic; routers decide which model should answer.