A new data format called TOON aims to dramatically reduce OpenAI API costs. Experiments with the format hint at future models like gpt-5-nano.
The OpenAI API has become a cornerstone for developers building a new generation of AI-powered tools, from chatbots and coding assistants to complex data-insight applications. However, a significant operational cost has always been the number of tokens processed, particularly when using verbose data formats. Now, a new, early-stage format called TOON (Token Oriented Object Notation) is emerging, promising to drastically reduce token usage and make interactions with models via the API significantly more efficient and affordable.
The High Cost of Conversation
For developers leveraging the OpenAI API, every piece of structured data sent to a large language model (LLM) — be it for prompts, function calls, or context — incurs a cost based on tokens. The standard format for this data interchange has long been JSON (JavaScript Object Notation). While ubiquitous and human-readable, JSON is not optimized for LLMs. Its reliance on braces, brackets, and repeated keys in large datasets leads to high token counts, which directly translates to higher operational costs and increased latency, as noted in a recent performance comparison.
TOON: A More Efficient Format Built for LLMs
TOON is designed specifically to address the shortcomings of JSON in AI workflows. By using an indentation-based structure instead of braces and a tabular notation for arrays of similar objects, it minimizes punctuation and redundant tokens. According to documentation on the new format, this token-centric design allows developers to send the same amount of information in a much more compact package. The goal is a lossless, predictable, and LLM-friendly serialization format that could significantly lower the barrier to building AI-powered products, aligning with OpenAI’s original mission for its API.
Glimpses of Future Models Emerge
Intriguingly, experiments and documentation surrounding these new efficiency-focused developments have revealed the names of unannounced OpenAI models. A practical experiment comparing TOON and JSON performance was conducted using a model identified as “gpt-5-nano.” Separately, OpenAI’s own documentation for image and vision capabilities features code examples that call a model named “gpt-4.1-mini.” While no official announcements have been made, these references suggest that a new wave of more specialized or efficient models may be in development, potentially being optimized to work with more compact data formats like TOON.
#OpenAIAPI #TOONformat #gpt-5-nano #LLMoptimization #AIdevelopment
