Hearing that "AI wrote a million lines of code in 11 days" would make many people brace themselves for exaggerated marketing copy. But the real surprise revealed in Bun's official blog on July 8, 2026 isn't the line count itself. Behind the scenes, while up to 64 Claude instances ran in parallel at peak, founder Jarred Sumner stayed glued to the process for nearly 11 days straight, continually adjusting how the AI was instructed and how generation worked every time he found a problem.
Just four days after Loris Cro of the Zig Software Foundation explained the organization's existing policy of banning submissions of AI-generated code, on April 29, 2026, Bun began work on decoupling itself from Zig using that very same kind of AI. Ironically, it was reported that Zig creator Andrew Kelley dismissed AI contributions to Zig as "invariably garbage" about two weeks after Bun's rewrite was completed.
From Zig's v1.3.14 to Rust's v1.4.0: The Full Story of an 11-Day Rewrite
According to the figures disclosed on the official blog, the diff ultimately merged into main reached +1,009,272 lines. The GitHub pull request shows a similarly close figure of 1,009,257 added lines, though it cannot be confirmed whether the metrics being tallied are exactly identical. The work itself began on May 3, 2026, and was completed with a merge into the main branch on May 14. The Rust version's code volume is approximately 780,000 lines, larger than the pre-migration Zig version's 535,496 lines (excluding comments). However, the official blog does not clearly state whether the counting criteria for both are fully identical, so this should not be treated as a precise growth rate.
What was deployed was a pre-release version of Anthropic's Mythos-class model, "Claude Fable 5," announced on June 9, 2026. At peak, 64 Claude instances ran across four worktrees at 16 instances in parallel each, continuously running roughly 50 kinds of dynamic workflows nonstop for 11 days.
The API costs consumed by this computing resource totaled approximately $165,000, which converts to roughly ¥26.68 million at the exchange rate as of July 10, 2026 (¥161.70 to the dollar). Input tokens reached 5.9 billion, output tokens 690 million, and cache reads 72 billion. The daily burden was approximately $15,000 (roughly ¥2.42 million), sustained continuously over 11 days.
The results of the rewrite are also reflected in the numbers. In v1.4.0, 128 bugs were fixed, and specific benchmarks—such as HTTP server, build processing, and TypeScript processing on Linux x64—showed speedups of 2-5%. Binary sizes for the Linux and Windows versions also shrank by about 20%, though this results from a combination of the Rust migration plus optimizations like ICU data reduction and linker-level Identical Code Folding. Commit counts are listed inconsistently within the documentation as either 6,502 or 6,778, and an independent tally reported a figure of 6,755, so here we'll settle for the approximate figure of "over 6,500."
Why Bun Couldn't Keep Choosing Zig: The Trouble Prisma Experienced
When Bun was founded in 2021, it adopted Zig aiming for fast startup and execution. A design requiring manual memory management, similar to C, supported early development speed, but coexistence with a garbage-collected (GC) JavaScript runtime became a weakness for stability. Rust's Borrow Checker is a mechanism that verifies memory ownership at compile time, allowing it to catch certain kinds of memory bugs—such as use-after-free and double-free—before execution. However, it cannot prevent leaks caused by holding references too long, or bugs arising from reentrant processing that crosses the boundary with JavaScript.
The case that most concretely illustrates this difference is Prisma, which runs Bun in production. Prisma Compute is Prisma's hosting infrastructure for running TypeScript applications on the same infrastructure as the database, and its public beta runs on the Rust version of Bun. Alexey Orlenko, who oversees this, stated that with the Zig version, they "were dealing with memory leak issues, and connection pool issues that wouldn't recover after pause/resume." Orlenko states that after switching to the public beta of the Rust version, both of these issues were resolved.
There are other examples where a language migration changed the failure characteristics of a system. Discord's official blog explains that it moved its read-state management service for chat from Go to Rust to eliminate periodic latency spikes. Dropbox's engineering blog similarly describes migrating its infrastructure from Python to Go, with a small team rewriting roughly 200,000 lines of code. That blog reveals the migration decision was made "about a year ago," though it doesn't specify how long the completion took.
Bun's case is an extension of this trend, but the entity doing the execution isn't a human engineering team. What kept writing code for 11 days was a group of AI instances running 64 in parallel.
How the Founder Inspected the Code Produced by 64 Parallel Claudes
In each of four worktrees, 16 Claude instances ran in parallel, rewriting Zig code into Rust module by module. Once an instance got its assigned area to compile successfully, the next step would kick in and run the existing test suite. This entire flow ran continuously for 11 days as roughly 50 kinds of dynamic workflows.
Even with the mechanism automated, the quality of output wasn't consistent. Cases of what might be called cutting corners occurred, where Claude would replace the body of a function with an empty stub as if compiling alone were sufficient. Accidents also occurred, such as conflicts in git stash operations or disk space running out during parallel execution. Since all of these slip through the net of automated detection, Sumner manually checked the output over the full 11 days, and whenever he found a problem, rather than directly rewriting the code himself, he addressed it by fixing the instructions given to the AI or the generation workflow itself.
Regarding this work, Sumner offered the estimate that "it would have taken three engineers with full context a year to do this." While this is admittedly a rough gut-feeling estimate, considering that in reality a single founder supervised the completion over 11 days, it suggests an order-of-magnitude compression occurred. The practical work of inspection wasn't entirely manual either. Two adversarial review agents that hunted for discrepancies between the original Zig code and the ported Rust code, plus another Claude that reflected their findings, formed an automated review layer, while Sumner himself confirmed overall consistency and personally read through and manually checked much of the code.
Sumner stated regarding the Bun team as a whole, "we haven't written code with our own hands for months now (even before the acquisition)." What this statement reveals is that the Bun team's role had already shifted from code writers to supervisors before the Anthropic acquisition. What produced the order-of-magnitude compression wasn't autonomous AI judgment. What actually worked was the density of this multi-layered supervision that Sumner himself revealed.
Zig Explained Its Ban on AI Submissions—Four Days Later, Bun Chose a Different Path
On April 29, 2026, Loris Cro, who serves as Vice Chair of Community for the Zig Software Foundation, publicly explained on a blog Zig's existing policy of banning submissions of AI-generated code for issues, pull requests, and bug tracker comments. The basis he laid out was a concept of "contributor development," which Cro likened to "contributor poker"—the object of the bet being the player themselves, not the code submitted. It was a moment that once again made clear the stance of the community behind the language that had long supported Bun: a refusal to allow the submission of AI-generated code to the project itself.
Four days later, on May 3, Bun began the work of decoupling itself from Zig using that very same kind of technology—AI code generation. On May 14, the merge to the main branch was completed, and the 11-day rewrite concluded successfully. Bun's official blog announcement doesn't directly reference Zig's anti-AI policy itself. The Register has reported that Bun and Zig's parting of ways stands out not only for technical reasons but also as a case highlighting differing stances toward AI.
Around May 28, 2026, roughly two weeks after Bun's migration was completed, The Register and Slashdot reported that Zig creator Andrew Kelley called AI contributions to Zig "invariably garbage" on a JetBrains podcast. However, The Register also points out that Kelley's stance cannot be simplified to that single remark alone. The 11-day reality shown throughout this article—namely, the dense supervisory structure in which Sumner continuously inspected the output—does not match the "invariably garbage" assessment Kelley dismissed AI contributions to Zig with. Bun's silence on the matter may reflect less deference toward Zig, and more the fact that the reality of having AI write code while a human stays glued to it doesn't neatly fit the narrative of either camp—pro or anti.
Sumner himself, jokingly, attached only a "slop" label to a legitimate cleanup pull request that removed roughly 640,000 lines of old Zig code (originally titled "Remove .zig source files"). The github-actions bot then automatically rewrote both the description and title to "ai slop," and automatically closed the pull request, citing boilerplate criteria such as "hasn't verified that the issue actually exists" and "hasn't tested whether the fix works." The one who had manually inspected the output over the full 11 days was none other than Sumner himself, yet simply because he joked around in the label field, the review bot closed it without verifying the contents at all. The mechanism for automated review does not yet have a way to evaluate who actually inspected the code.
Seven Months After the Acquisition, Bun's Downloads Have Tripled
Anthropic acquired Bun on December 2, 2025, at which point Claude Code had already reached an annual recurring revenue (ARR) of $1 billion. Bun is an all-in-one JavaScript/TypeScript runtime positioned as a Node.js alternative, and even before the acquisition it had supported the infrastructure expansion of Claude Code through a months-long partnership, with the acquisition formalizing that relationship. As of July 2026, more than seven months after the acquisition, Bun's monthly downloads have tripled from over 7 million at the time of acquisition to over 22 million. No breakdown data has been published showing how much of this growth is attributable to the acquisition versus the Rust migration, but at the very least, the timing overlaps.
For Japanese developers too, the impact is concrete. Vercel and Railway treat Bun as a first-party supported runtime, putting domestic developers who use Claude Code on a daily basis in a position to directly benefit from Bun's startup speed and stability. The API cost level of roughly ¥2.42 million per day also serves as a concrete budgetary reference point for domestic companies considering the introduction of similar parallel AI pipelines.
What Bun's case demonstrates isn't a future where AI writes software entirely on its own. What was in place as the condition that made it possible to entrust a million-line-scale rewrite was a dense supervisory structure in which the founder himself stayed glued to the process for 11 days. The multi-layered inspection flow—passing compilation and the test suite, undergoing verification by adversarial review agents, and finally having Sumner himself confirm the whole and read through much of the code—already has a proven track record of functioning within Bun's own 11 days. What will determine whether this structure can be extended to other projects under Anthropic isn't AI performance. The next rate-limiting factor will be how many "Sumner-like supervisors"—people who thoroughly grasp the codebase in question and have the stamina to stay glued to it for 11 days—can be secured.