March 26, 2026
JSON beef, well done
We Rewrote JSONata with AI in a Day, Saved $500K/Year
AI nukes a $500K bill; commenters ask how it was that pricey and why not use what already exists
TLDR: A startup used AI to quickly rebuild a data tool and claims it slashed about $500K a year in costs. Commenters are split between applause and disbelief, questioning the huge original bill, why existing tools weren’t used, and whether they should dump JSON for faster binary formats—making this a wallet win turned culture war.
An engineer pulled a weekend heist with AI: rebuild a data‑shaping tool in Go, pass the tests, claim a 1,000x speedup, and boom—$500K a year saved. They say it took seven hours, about $400 in AI tokens, and 13,000 lines of code. Cue the internet dogpile. The crowd is split between slow clap and side‑eye. One camp is cheering the “spec + tests + AI” playbook (inspired by Cloudflare’s AI Next.js rewrite) and the instant cost drop. The other camp can’t get over the sticker shock of the old setup and the idea of shipping tiny data checks across the network just to get an answer. As one commenter put it, “I cannot fathom this level of cost.”
Then the debate turns spicy. Practical folks ask why they didn’t use existing Go versions listed in the official JSONata docs. Another voice throws shade by noting the JavaScript reference is only 5.5k lines—so 13k in Go raised eyebrows. Meanwhile, the performance purists are yelling, “Forget JSON—go binary!” and dropping cheeky tips about not repeating field names in every message. Memes fly about “AI firing 200 pods” and “microseconds vs nanoseconds couples therapy.” It’s classic tech internet: a feel‑good save‑the-money story instantly transformed into a cage match over costs, choices, and whether plain old JSON is the real villain.
Key Points
- •Reco replaced an RPC-based jsonata-js setup with a pure-Go JSONata 2.x evaluator called gnata.
- •Using an AI-and-tests approach, gnata was built in ~7 hours for ~$400, achieving 1,778 passing tests across ~13,000 lines of Go.
- •The prior architecture incurred ~$300K/year compute costs and significant latency (≈150µs per RPC round-trip), with large Kubernetes deployments.
- •gnata delivered up to a 1,000x speedup on common expressions and reduced a $25K/month jsonata-js production cost to zero.
- •gnata’s two-tier evaluation classifies expressions at compile time and fast-paths simple cases and 21 built-in functions; features include streaming focus, localized caching, WASM support, metrics, and fallthrough to jsonata-js RPC.