Verbosity and Redundancy

JSON's syntax is inherently verbose. It uses a lot of punctuation like curly braces, brackets, and quotes. Keys are repeated for every object in an array, leading to significant token waste.

[
  {"id": 1, "name": "Alice"},
  {"id": 2, "name": "Bob"}
]

The keys "id" and "name" are repeated, and all the punctuation adds to the token count.

Lack of Schema Enforcement

JSON has no built-in schema. You need to describe the desired structure and data types within the prompt itself, consuming valuable context window space and increasing the chances of the LLM generating non-compliant output.

Syntactic Rigidity

A single misplaced comma or quote can render an entire JSON document invalid. LLMs can easily make these small mistakes, requiring developers to implement robust validation and parsing logic to handle errors.

{"name": "Alice", "age": 30,}

The trailing comma after '30' makes this JSON invalid in many parsers.

Inefficient Tokenization

LLM tokenizers often split punctuation and special characters into multiple tokens. This means characters like '{', '}', '"' can consume more tokens than a single character, further inflating costs.