Anthropic Migrates Bun's 1M Lines of Zig to Rust in 11 Days for $165k: The Real Cost Breakdown

The $165k Migration That Reset Expectations

Anthropic published a case study on July 16, 2026 describing how Jarred Sumner, creator of the Bun JavaScript runtime, used Claude Fable 5 to migrate roughly 535,000 lines of Zig into Rust in 11 days. Total Anthropic API bill: about $165,000. Sumner orchestrated 64 parallel agents through Claude Code, each handling a slice of the codebase with tests as verification gates.

On its face this is a marketing story for Fable 5. But the numbers tell a more useful story for anyone budgeting AI-assisted work: cost per line, cost per parallel agent-hour, and the specific conditions that made this project tractable.

Unit Economics of the Migration

The reported numbers break down to about $0.31 per line of source migrated. That's below what a mid-level engineer costs even in low-cost regions, and it's an order of magnitude below what traditional consultancies quote for language migrations ($3-$8 per line historically). But three things made this possible: existing tests, an owner with deep architectural knowledge, and a model priced for the workload.

Metric	Value
Source lines migrated	~535,000
Total API spend	~$165,000
Cost per line of code	~$0.31
Wall clock time	11 days
Parallel agents	64
Avg agent-hours per day	~1,536

Model Choice: Fable 5 vs the Alternatives

Fable 5 sits at $10 input and $50 output per million tokens. That's expensive per token, but the case study makes clear the value was in reducing retry loops — Fable 5's higher first-pass correctness meant fewer wasted agent runs, and 64 parallel agents amplify any per-run efficiency gain. Here is what the same migration would have cost using different frontier models, assuming rough parity in per-line efficiency (which is optimistic for cheaper models):

Model	Input $/M	Output $/M	Naive scaled cost
Claude Fable 5 (actual)	$10	$50	$165k
GPT-5.5 Pro	$30	$180	$530k
GPT-5.6 Sol	$5	$30	$95k
Claude Opus 4.8	$5	$25	$82k
Gemini 2.5 Pro	$1.25	$10	$28k
DeepSeek V4 Pro	$0.435	$0.87	$3.5k

The "naive scaled cost" column is what you'd pay if token counts held constant across models — which they never do. Cheaper models typically require 2-4x more retry loops for the same first-pass correctness on a task this complex, which erodes the paper savings quickly. The DeepSeek V4 Pro number at $3.5k looks tempting but assumes zero rework, which is unrealistic for a 535k-line Zig-to-Rust migration.

What Made This Migration Cheap

Three preconditions made $0.31 per line achievable, and if your project is missing any of them your actual cost will look nothing like Sumner's:

1. Extensive existing test coverage. Bun's tests acted as the acceptance oracle for each migrated slice. Without that, human review time would have dominated and blown the budget past 5-10x.

2. An owner with architectural clarity. Sumner knows Bun end-to-end. He wrote the original Zig. That let him decompose the work into 64 independently verifiable slices — a task the AI itself would have struggled with cold.

3. Static, well-defined transformation. Zig-to-Rust is a mechanical rewrite of memory management patterns, not a redesign. Migrations that change architecture — say, monolith to microservices — do not exhibit this economics.

Should You Copy This Playbook?

For a language migration on a codebase with strong tests and an engaged owner, yes. Budget $0.30-$1.00 per line of source depending on complexity and add 25-40% buffer for review overhead. For architectural rewrites, greenfield features, or anything requiring new design decisions, this economic model does not transfer.

Use our AI cost estimator with Fable 5 or Opus 4.8 as the model and "large refactor" as the project type to model your own migration. Compare against DeepSeek V4 Pro and Kimi K2.7-Code — those two are increasingly the smart mid-tier choices for teams that can tolerate slightly more retries in exchange for a 10-20x lower per-token cost.