Bun's Rust Rewrite: 11 Weeks, $165,000, and No Release in Sight

Bun's Rust Rewrite — The Reality of AI Coding

In May 2026, Jarred Sumner, creator of the Bun JavaScript runtime, announced what seemed like a watershed moment for AI-assisted software development: Bun had been rewritten in Rust using Anthropic's Claude AI. The entire process took just 11 days and cost $165,000 in API calls. Tech media erupted with celebration — finally, proof that AI could do the work of open-source maintainers, faster and at scale.

But as Tom Lockwood's meticulous investigation reveals, the reality is far more complicated than the breathless headlines suggested. Eleven weeks after the last official Bun release (v1.3.14, May 12, 2026), there is still no new release tag. The rewrite was merged to main, but six weeks later, not a single release has been cut.

The Investigation: What Lockwood Found

Lockwood took the simple but effective step of cloning the Bun repository — all 1.23 GB of it — and examining the release tags firsthand. The results were striking. As of July 27, 2026, the git log showed:

2026-05-12 15:12:49 -0700  (tag: bun-v1.3.14)

This 11-week gap is unprecedented in Bun's recent history. The last time the project went over a month without a release was back in October-December 2022, between v0.2.2 and v0.3.0 — an entirely different era for the project.

What's happening under the hood is even more revealing. On July 9, robobun (the bot proxy for Claude Code-generated pull requests) had 1,277 open PRs. By July 27, that number had ballooned to 2,475 open PRs. The project's Buildkite CI/CD cluster has been running continuously since the "rewrite" was "completed" — Lockwood's scare quotes are deliberate. At the current rate, merging all open Claude-generated PRs would take approximately 86 days of continuous pipeline operation.

Some PRs take 40 minutes to an hour and a half to clear Buildkite checks. Many of them deal not just with Rust code, but with the complex web of dependencies and configurations that make up a production JavaScript runtime.

The Real Cost: Far Beyond $165,000

Lockwood's analysis reveals that the widely-cited $165,000 figure represents only a fraction of the true cost. Here's what's not included in that number:

• CI/CD Infrastructure: The Buildkite cluster's constant operation since May represents a significant — and undisclosed — expense.
• Anthropic Employee Involvement: Analysis of commit patterns shows Anthropic employees directly contributing Rust code and reviewing PRs. Their salaries are not reflected in the API cost.
• Ongoing Claude Usage: The project continues to burn through Claude credits at a substantial rate. Lockwood estimates that if ongoing costs are $10,000 per day, the total spend is approaching $800,000.
• Unaccounted Credits: Some PRs show signs of additional Claude credits being applied beyond the initial $165,000 allocation.

The $15,000-per-day API burn rate during the rewrite is — as Lockwood puts it — "well outside the means of many open source maintainers." For context, that's more than most independent developers earn in a month, spent daily on API calls alone.

A Pattern of AI Hype

Bun's rewrite is not an isolated case. Lockwood draws parallels to two other high-profile AI-generated projects:

• Anthropic's C Compiler: A Claude-generated C compiler that garnered significant attention — and then went completely silent. No commits for months.
• Cursor's FastRender Browser: An AI-generated web browser that similarly captured headlines — and similarly stopped receiving updates.

This pattern raises uncomfortable questions about the AI coding narrative. Are these projects genuine breakthroughs, or are they marketing exercises designed to support massive company valuations? As Lockwood puts it: "I think it's important to be very Canny when someone makes a claim that supports a company's large valuation."

Anthropic's acquisition of Bun before the rewrite adds another layer of context. The rewrite serves as dogfooding — Anthropic using its own product to build something impressive, while simultaneously generating proof-of-concept marketing material. There's nothing inherently wrong with this, but it means the claims should be examined with appropriate skepticism.

What This Means for AI Coding

Lockwood's investigation doesn't dismiss AI coding outright — far from it. He acknowledges his own ML/AI work as some of his most proud accomplishments. But he makes a crucial distinction: AI as a tool versus AI as a replacement for engineering discipline.

Several key lessons emerge for developers evaluating AI coding tools:

1. Generation is not completion. Writing code is only a small part of software engineering. Testing, debugging, optimization, security auditing, and integration all remain human-intensive tasks — regardless of who (or what) wrote the initial code.

2. Hidden costs matter. The headline $165,000 figure doesn't capture the full cost of production-ready AI-generated code. CI/CD time, human review, ongoing API costs, and infrastructure all add up — sometimes to multiples of the initial investment.

3. Maintenance is forever. Even if AI can write code faster, it still needs to be maintained. The silence around Anthropic's C compiler and Cursor's FastRender suggests that AI-generated projects may have a shorter maintenance lifespan than traditionally-built ones.

4. Context is everything. The Bun rewrite happened within a specific context: an Anthropic-acquired company using an Anthropic product, with Anthropic employees directly involved, backed by significant venture capital. This is not replicable by most open-source projects or small teams.

The Sustainability Question

Perhaps the most fundamental question Lockwood raises is one of sustainability. If a well-funded, VC-backed project with direct access to the AI company's employees is struggling to ship a release six weeks after the "completed" rewrite, what hope does the average open-source maintainer have?

The answer, for now, seems to be: AI coding tools are powerful accelerators, but they are not yet replacements for traditional engineering practices. The $165,000 figure that made headlines is better understood as a marketing cost than a production cost — the price of demonstrating what's possible under ideal conditions, not what's practical for everyday development.

Conclusion

Lockwood ends his investigation with a simple, devastating question: "Are we done yet?" The answer, for Bun and for AI coding generally, appears to be no. Not yet. The rewrite may eventually ship, and AI coding may eventually live up to its promise. But as of July 2026, with no release in sight and CI/CD clusters running 24/7, the evidence suggests that we are still in the hype phase — not the delivery phase.

For developers evaluating AI tools, the lesson is clear: be enthusiastic about the potential, but be Canny about the claims. Clone the repo. Check the tags. Look at the CI pipeline. The truth is in the code, not in the blog posts.

📖 Source