Hacker News front pageJun 13, 2026, 04:33 UTC+8

A cross-vendor agent loop: Claude Fable 5 as architect, GPT-5.5 Codex as builder

/architect: Reduce Fable tokens by 80%, Fable orchestrates/reviews, Codex builds

Dan McInerney open-sourced a Claude Code skill that chains Claude Fable 5 and GPT-5.5 Codex into a division-of-labor loop. Claude plans and reviews, Codex writes code, and the repo acts as memory. The author claims an 80% reduction in Fable token usage, but the post doesn't include benchmarks or comparison data—just the README and code, so real-world results are unverified.

Why it matters: A runnable cross-model agent loop with a concrete 80% token-saving claim. Claude-as-architect + GPT-as-builder is a practical pattern worth testing. Score held at 72 because no benchmarks or third-party validation are provided — it's all self-reported.

Read the original ↗Export Markdown

Take

This caught my eye because it turns cross-model division of labor into a reusable Claude Code skill. Dan McInerney has Claude Fable 5 act as architect—breaking down tasks and reviewing code—while GPT-5.5 Codex does the actual building, with the repo serving as shared memory. He claims this cuts Fable token usage by 80%, but the repo only has a README and code, no benchmarks or comparison data.

I'd discount that 80% figure. It reads like personal experience, not a systematic eval. The cost logic makes sense—Fable 5 isn't cheap, and offloading most calls to Codex would save money—but only if Claude's planning and review are solid enough to avoid rework loops from lost context. That's exactly what we don't have data on.

Treat this as a cost-saving template for narrow tasks, not a general solution. It probably works best when the task is well-defined, Codex can handle it independently, and Claude only needs light review. For complex refactors or tasks requiring deep codebase understanding, this split might add more back-and-forth than it saves.

Source: Hacker News front page · github.com
Published: Jun 13, 2026, 04:33 UTC+8 · Jun 13, 2026