Why JSON wastes LLM tokens, and the lossless fix

Every time you pass structured data to a language model, you pay twice for JSON's design decisions: once in dollars, once in context window.

Where the waste comes from

JSON repeats every key for every object in an array. It wraps everything in braces, brackets and quote marks. Tokenisers dutifully split all of that into tokens that carry no information the model actually needs — the hundredth repetition of a field name tells the model nothing the first one did not.

In a single API call this is a rounding error. In agentic systems it is not. Agents pass structured data through the model many times per task: tool results, state, intermediate outputs. The overhead compounds with every hop, and long conversations hit the context ceiling earlier than they should.

What XRON does

XRON is a lossless data serialisation format I wrote in TypeScript. Lossless is the operative word — nothing is summarised, truncated or approximated. The same information is restructured to consume up to 80 per cent fewer tokens.

The result is practical, not academic: context windows stretch further, GPT and Claude costs drop, and multi-agent systems can run longer conversations before degrading.

Why a marketer built a serialisation format

Performance media taught me two habits: automate the repetitive work, and obsess over data quality, because every downstream decision inherits it. LLM workflows over structured data — feeds, reports, campaign objects — have exactly the same shape. The format of your data is an input cost like any other, and it was being ignored.

XRON is open source on GitHub. The cheapest token is the one you never send.