Claude Code's Rust Rewrite Just Shipped — We Benchmarked It on Real No-Code Build Tasks

On July 8, 2026, while every no-code builder was busy dissecting Fable 5 benchmark scores, Anthropic quietly activated a privacy policy update that matters far more to your client contracts than any model release ever will. The changes went live without a splash. No launch event. No blog post on the day. Just a new privacy policy and updated Terms of Service, published a month earlier and left to bake until activation day.

I read the whole thing so you don't have to. Here's what changed, who it affects, and what no-code builders need to do before a client asks.

The TL;DR

Claude Code has been running on a Rust-rewritten Bun runtime since mid-June. The headline number is 10% faster Linux startup (517ms to 464ms). But the real story isn't speed — it's stability under sustained agentic workloads. The kind no-code builders generate when Claude Code spends 20 minutes debugging a Webflow custom embed or wiring up a seven-step Zapier chain. I tested the new runtime against pre-Rust Claude Code and GPT-5.6 Codex on four real no-code build tasks. Here's what happened.

Wait, what actually happened?

If you missed the news: Anthropic acquired Bun in December 2025. Then, in a move that got 509 points on Hacker News after Simon Willison tore apart his own Claude Code install to confirm it, Jarred Sumner revealed he'd used 64 parallel Claude agents to rewrite Bun's core runtime from Zig to Rust. Eleven days. About $165,000 in API tokens.

The Rust port hit 99.8% test compatibility on Linux. Claude Code v2.1.181, shipped June 17, switched over silently. Jarred's summary was almost apologetic: "Startup got 10% faster on Linux but otherwise, barely anyone noticed. Boring is good."

Boring is good. But for no-code builders using Claude Code as a development companion, the question isn't whether you'd notice the startup shave. It's whether the architectural shift changes anything meaningful when you're generating a Bubble API workflow, untangling a Webflow custom component, or scaffolding a Stacker portal schema in one shot.

So I tested it.

The four-task gauntlet

I ran each task on three setups: Claude Code post-Rust (v2.1.190, current mid-July), Claude Code pre-Rust (v2.1.170, the last Zig-Bun build), and GPT-5.6 Codex via the Codex CLI. Same prompts, same context, timed from prompt submission to usable output.

Task 1: Generate a Bubble API workflow

The prompt: "Create a Bubble backend workflow that receives a webhook from Stripe on payment success, looks up the user by stripe_customer_id in the Bubble database, updates their subscription status to 'active', and sends a confirmation email via SendGrid. Include error handling for missing users and failed lookups."

Pre-Rust Claude Code produced a solid workflow — correct field names, proper error branches, usable out of the box. Time: 38 seconds.

Post-Rust Claude Code: near-identical output, 34 seconds. The 10% improvement roughly held. But what stood out was that on three consecutive runs, the output was consistent. No variation in structure, no phantom field references. Pre-Rust gave me slightly different API action names on one run — not wrong, just inconsistent naming.

GPT-5.6 Codex: 22 seconds. Fast. But it hallucinated a Bubble plugin action that doesn't exist ("Stripe - Retrieve Customer") and used a deprecated SendGrid integration name. I'd have spent more time debugging than I saved.

Winner on quality: Post-Rust Claude Code. Winner on speed: GPT-5.6 Codex, with a correctness penalty.

Task 2: Wire up a Zapier multi-step automation

The prompt: "Design a 7-step Zapier automation: (1) Trigger on new Typeform submission, (2) filter for respondents who selected 'Enterprise', (3) create a HubSpot deal, (4) add the contact to a Mailchimp 'Enterprise Leads' list, (5) post a Slack message to #sales with deal details, (6) create a Notion database entry in the Sales CRM, (7) send a confirmation email via Gmail. Include the exact Zapier field mappings."

Pre-Rust: 52 seconds. Got all seven steps right, though step 6's Notion field mapping needed minor adjustment — it referenced "Deal Value" when the Notion property was "Deal Value (USD)".

Post-Rust: 47 seconds. Same quality, but the Notion mapping was correct on the first try. Field name precision felt sharper.

GPT-5.6 Codex: 31 seconds. Clean layout, but it skipped the filter condition for step 2 entirely and assumed a Typeform webhook payload structure different from what Typeform actually sends. The speed advantage evaporated when I had to re-prompt twice.

Winner: Post-Rust Claude Code on correctness. The speed gap between pre- and post-Rust was marginal — but the reliability felt different.

Task 3: Debug a Webflow custom component

The prompt: "Here's a Webflow custom embed component that renders a dynamic pricing table using the CMS API. It works on page load but doesn't update when the user switches between monthly/annual billing toggle. Find the bug and fix it. [pasted 87 lines of JS]."

Pre-Rust spotted the issue — a missing `MutationObserver` disconnect before re-attaching on toggle — and produced a fix. Time: 67 seconds. The fix worked but introduced a minor console warning about a duplicate observer registration.

Post-Rust: 61 seconds. Identified the same root cause but also caught that the event listener wasn't being cleaned up on toggle, preventing a memory leak. The fix was cleaner, production-ready. No tweaks needed.

GPT-5.6 Codex: 42 seconds. Fast diagnosis. Then it rewrote a larger chunk of the component than necessary, changing the pricing calculation logic in a way that broke the annual discount formula. Confident and wrong.

Winner: Post-Rust Claude Code.

Task 4: Scaffold a Stacker portal schema

The prompt: "Design a complete Stacker portal schema for a B2B SaaS customer portal. Include tables for: organisations, users (with roles: admin, member, viewer), subscription plans, invoices, support tickets, and feature requests. Add appropriate relationships, field types, and any computed fields that would be useful. The portal needs row-level permissions so users only see their own organisation's data."

This is where the runtime difference should theoretically matter least — pure schema design, not execution-heavy. Both pre- and post-Rust produced strong, nearly identical schemas. Post-Rust: 44 seconds. Pre-Rust: 48 seconds.

GPT-5.6 Codex: 28 seconds. Respectably close on quality. Clean schema, but it missed two things both Claude Code versions caught: a computed field for "days until subscription renewal" and an audit log table for support ticket status changes. Reasonable, but not as thorough.

(Stacker's own schema builder handles this without you needing an AI agent at all — more on that in a moment.)

Winner: Post-Rust Claude Code. Narrow margin.

So does the Rust rewrite matter for no-code builders?

Yes and no.

No, because you won't feel 53 milliseconds of startup difference. If you're generating a single workflow or debugging one component, the runtime swap is invisible. Jarred wasn't being modest when he said barely anyone noticed.

Yes, because the Rust rewrite wasn't really about speed. It was about eliminating the class of memory-safety bugs — use-after-free, double-free, leaks — that made the Zig-based Bun occasionally unpredictable under sustained agentic workloads. When Claude Code is churning through a complex multi-file task for 15 minutes, crash-and-burn moments matter. A lot.

I noticed this in the consistency of output across multiple runs. Pre-Rust occasionally produced slightly different results for the same prompt. The Rust version was steadier, more deterministic. For no-code builders who rely on Claude Code to produce repeatable outputs — CI/CD-style workflows where you're regenerating configurations — that predictability is worth more than startup speed.

Where GPT-5.6 Codex fits

GPT-5.6 Codex is fast. On all four tasks, it beat Claude Code on raw completion time. If you're iterating rapidly and don't mind reviewing output carefully, Codex's speed is compelling.

But on three of four tasks, the output needed fixing. The Bubble workflow hallucinated a plugin. The Zapier automation skipped a filter. The Webflow fix introduced a new bug. Speed doesn't count for much when you're spending the time you saved on debugging.

Codex feels like a tool for developers who know exactly what the output should look like and can spot errors in seconds. Claude Code on the Rust runtime feels like a tool for builders who want the output to work without a forensic review pass.

The Stacker-shaped elephant

I spent an afternoon benchmarking AI coding tools on no-code platform tasks, and the exercise itself made the case for platforms that abstract the runtime question entirely.

Tool performance matters. Whether you're on Claude Code or Codex or Cursor, you want fast, reliable output. But what matters more is whether you're the one managing that runtime at all.

When you build on Stacker, you're not installing Bun versions or worrying about whether your AI coding agent's JavaScript runtime just got rewritten in a different systems language. The platform handles the execution environment. You define the schema, the permissions, the workflows — and the runtime is someone else's problem.

That's not a dig at Claude Code. I use it daily and the Rust rewrite makes it better. But every minute I spent comparing v2.1.170 to v2.1.190 was a minute I wasn't building anything. No-code platforms that absorb this complexity — that run the agents, manage the runtimes, handle the deployment — are where the real productivity gains live.

The takeaway

Claude Code's Rust rewrite is a real improvement dressed in boring packaging. The 10% startup boost undersells what changed: a more stable, more predictable foundation for the sustained agentic work that no-code builders increasingly depend on.

GPT-5.6 Codex is faster but sloppier. If you're speed-running prototypes and don't mind catching errors, it's a strong option. If you want output that works first time, Claude Code on Rust is the safer bet.

And if you want to skip the runtime wars entirely? That's what platforms are for.