November 2025
bear-1: First LLM Input Compression Model
bear-1 is a compression model that reduces LLM input tokens by 66% without sacrificing accuracy. It integrates as a single API call before your existing LLM pipeline.
What is bear-1?
bear-1 is The Token Company's first LLM input compression model. It uses semantic compression to intelligently remove redundant tokens from prompts before they reach your language model. The result: you send fewer tokens to your LLM, pay less, and get the same (or better) results.
Unlike simple truncation or summarization, bear-1 understands the semantic structure of your input. It identifies and removes tokens that don't contribute to the meaning of the text — what we call “context bloat” — while preserving the information your LLM needs to generate accurate responses.
Performance
66%
Token reduction
3x
Cost reduction
+1.1%
Accuracy gain
<0.1s
Per 10K tokens
Benchmarked on GPT-4o-mini using LongBench v2, bear-1 achieves a 66% reduction in input tokens while actually improving accuracy by up to 1.1%. This means you're not just saving money — your LLM performs better with compressed input because the noise has been removed.
How it works
Integration is a single API call. Send your text to the bear-1 compression endpoint, get compressed text back, and pass it to your LLM. That's the entire integration.
Pricing
$0.05 per 1M compressed tokens. You only pay for the tokens that were removed. The tokens that pass through are free.
What's next
bear-1 laid the foundation for LLM input compression. We've since released bear-1.1, an improved version with better accuracy preservation and faster compression. If you're starting fresh, we recommend bear-1.1.
Start compressing with bear-1
Create an account and get your API key.