TwoTrim AI
Safest way to save the AI token costs.
Start new thread
Ash M

TwoTrim AI - Make Every Token Count, Save on LLM API Bills

by
Maker Town
•
TwoTrim is a research-backed, lossless token compression engine that reduces LLM input tokens by 20–60% with no change in output quality. It is stateless, secure, and drop-in compatible with OpenAI, Anthropic, and Gemini. Built from a published study across 50,000+ prompts, TwoTrim delivers predictable savings with zero deployment cost and zero data storage. Our goal is simple: help AI builders scale without scaling their API bills.

Add a comment

Replies

Best
Ash M
Maker
📌
Hi all, Ash here from TwoTrim! 👋🏻 I wanted to share what exactly inspired us to build this? Over the last few months, we kept hearing the same thing from AI founders: “Our biggest bottleneck isn’t the model. It’s the API bill.” We decided to test whether LLM prompts could be compressed without losing meaning. What started as a side experiment turned into a full research paper evaluating 50,000+ prompts. The results surprised us. We consistently reduced input tokens by 20–60 percent while maintaining 100 percent output fidelity. So we built TwoTrim. A lossless, stateless, drop-in token optimizer that requires no code changes and stores zero data. We’re giving it away as free deployment model because we want to help AI teams ship faster, scale cheaper, and focus their resources on what matters: product. Would love your feedback, questions, or use cases. Happy to share more technical details too! 🚀
Ash M
Maker

Here’s our research summary 📘

We evaluated TwoTrim across 50,000 diverse prompts spanning reasoning, chat, structured tasks, analysis, and JSON workflows.
Some of the key findings:

  • 25.96% average token reduction with perfect output fidelity

  • 62.1% of prompts achieved 20–40% savings

  • Mathematically guaranteed invertibility, ensuring outputs remain identical to uncompressed prompts

  • Model-agnostic design, validated across OpenAI, Anthropic, Gemini and others

We built TwoTrim as a research project first, then turned it into a tool for builders once the numbers consistently held up.

Read the full paper here:
🔗 https://www.twotrim.com/resources/research

Happy to answer anything about methodology, benchmarks, or implementation!