Tokenmaxxing and the Collapse of Coding Costs

Tokenmaxxing and AI coding agents are accelerating software development by allowing individuals and small teams to delegate complex programming tasks to multiple AI systems, potentially reducing development costs and changing how software companies compete.

On March 31, 2026, Anthropic published what looked like a routine update to Claude Code, its fast-growing AI coding agent. At the time Claude Code’s annualized revenue was $2.5 billion.

Buried inside the update was something that was never supposed to leave Anthropic: A 59.8-megabyte source-map file. Inside it: roughly 512,000 lines of TypeScript spread across 1,900 files—a sprawling window into Claude Code’s internal architecture, including its agent orchestration, tools, system behaviour and experimental features. Anthropic had inadvertently published the code itself.

Security researcher Chaofan Shou discovered the mistake and posted about it on X (formerly Twitter) and included the download link to where Anthropic had published it. The second person to reply on X was Sigrid Jin. In less than 30 minutes  Jin had downloaded all the code.

Jin, whose legal name is Jin Hyung Park, is a Korean developer and University of British Columbia student who divides his time between Seoul and Vancouver.

The Wall Street Journal reported that he consumed 25 billion tokens in 2025. At Web Summit Vancouver in May 2026, Jin said that number was under reported, and that he’d actually burned through 50 billion tokens in 2025. Jin is the poster boy for tokenmaxxing. (A token is the basic unit of information processed by a large language model – think of it as a word.)

Most people have no idea what advanced AI systems are capable of because they aren't using them in a significantly way. Buying a $20-a-month AI subscription and occasionally asking it to draft an email, is like buying a supercomputer and using it as a calculator. Jin argues that serious AI users should spend as much on AI as they spend on rent.

When Claude Code’s source appeared online, Jin went to work. At the time he was on a plane from Seoul to Vancouver. Using the Wi-Fi, he downloaded the half million lines of code. And then working with Seoul-based developer Yeachan Heo, they used Codex to recreate Claude Code's core architecture in Python using the leaked TypeScript code, with a Rust version following later. They called it Claw Code.

On GitHub, Jin and Heo’s Claw Code repository was the fastest ever to reach 100,000 stars in less than 24 hours.

Jin described the effort as two humans, ten AI agents and a MacBook Pro. The work was orchestrated through a workflow built on OpenAI’s Codex. Instead of two programmers laboriously working through hundreds of thousands of lines of code, multiple AI agents could operate simultaneously. One could analyze a component. Another could implement it. Another could run tests. Another could inspect the first agent’s work. Another could compare the behaviour of the new system with the original. Then they could iterate. And iterate again. Jin and Heo weren’t programming; they were managing programmers that happened to be machines.

Software copyright has always rested on an important distinction. Copyright protects the particular expression embodied in source code. It does not automatically give a company ownership over every idea, function, behaviour or architectural concept contained within a program.

That distinction made sense in a world where independently recreating a sophisticated piece of software could take dozens of engineers’ months or years. AI has changed all that. Jin’s Python implementation appeared within hours. And later it was translated into Rust, a programming language prized for performance, memory safety and systems-level efficiency.

For decades, software companies have enjoyed some of the most attractive economics in business. Creating great software is expensive. Copying and distributing it costs almost nothing. The expensive part has always been the people required to design, build, test and maintain it. In other words, the scarce resource was engineering talent. But AI coding agents are beginning to erase some of that scarcity.

Jin’s extraordinary token consumption wasn't simply digital extravagance. Those billions of tokens represented an enormous volume of delegated work. Agents writing code. Testing it. Debugging it. Reading documentation. Refactoring programs. Reviewing one another’s work. Correcting mistakes. And trying again.

By the time Claude Code’s source became visible, Jin had spent more than a year developing a skill that barely existed a few years earlier: managing digital workers.

The human role moved upward—from producing every line of code to decomposing the problem, assigning work, evaluating outputs, exercising judgement and deciding what happened next. That is agentic software engineering. And it creates a new kind of leverage.

The traditional software organization might look something like this:

Manager → Engineering Manager → Team Lead → 10 Programmers

Increasingly, it may look more like:

One expert → AI orchestrator → 10, 50 or 100 specialized agents

The implication isn't necessarily that humans disappear. It is that where humans create value changes. Typing code is no longer important. Knowing what should be built, how to break the problem apart, how to judge the output and when the machine is wrong becomes more important. The programmer becomes a conductor.

The irony is that Claude Code was built to automate software development. Then Claude Code itself became the target of automated software development.

The marginal cost of creating code has plummeted. The speed at which software can be understood, reproduced, transformed and improved has shifted from years to hours. Proprietary code alone becomes a weaker competitive moat.

The things that matter most may shift toward assets that are harder to copy: proprietary data. Distribution. Customer relationships. Infrastructure. Brand. Trust. Ecosystems. Execution. And perhaps the ability to innovate faster than somebody else can reproduce what you built yesterday.

For most of the history of computing, software development was constrained by the reality that humans could write only so much code, so quickly. The Claude Code episode offers a glimpse of what happens when that constraint disappears.

Jim Harris is a #1 International Bestselling author published in 80 countries. His new book The Age of Fakes!: How AI Abuse, Fake News, and Deepfakes Threaten Business and Society is a #1 International bestseller in multiple categories and countries. You can see it at https://www.amazon.com/Age-Fakes-Deepfakes-Threaten-Business/dp/3982801001/ref=sr_1_1

Follow Jim on Linkedin.com; X (Formerly Twitter), YouTube and Substack

Interested in having Jim as an AI keynote speaker at an upcoming event? Email him at [email protected]