condense.chat

LLM input compression. We build the compression models. Drop-in proxy for Claude Code, Codex, and OpenCode. Cuts input tokens by up to 70%.

[condense.chat @densechat]

We're giving you 100M free tokens to prove your agent is wasting context.

Most of what your agent sends upstream is dead weight. We built state-of-the-art compaction to strip it, and today we published the receipts.

One real session, run to full depth, benchmarked against headroom:
condense: 53.8% of tokens saved, 37.3% off the bill
headroom: 15.1% and 13.8%

Everything is open in the post. One curl line, and triple your Fable usage.

Full breakdown in the comments

[Ed @edvard_exe]

everyone’s celebrating the new opus. nobody’s mentioning it quietly raised your bill the second you upgraded.

same code. same prompts. same work. but the new tokenizer counts about 34% more tokens. and you pay per token. it’s in anthropic’s own docs.

we built @densechat for exactly this. strip the tokens your agent no longer needs. your bill lands lower than before the new model existed.

new generation. old problem. already fixed.

new benchmark drops next week.

[condense.chat @densechat]

Our last team member: @edvard_exe

>27 years old.
>built his first profitable business at the age of 5.
>at school, hated coding and said he would never code.
>background in aerospace engineering.
>was building UAVs and morphing wings before it was cool.
>studied ML engineering for two years in Italy.
>spent years building various models.
>was shipping side projects the whole way through.
>built one of the best deep research engines.
>3x hackathon winner.
>cofounded https://t.co/gfokj5Errj, responsible for the BM25 implementation.

bullish on: world models

[condense.chat @densechat]

Two more to go!

Welcome: @packgron

>29 years old.
>started programming at the age of 14.
>taught himself linear algebra at 15.
>wrote his own game engine at the age of 16.
>started working as a software engineer right after school.
>polyglot in multiple programming languages, from assembly to JS.
>currently shilling Rust.
>implemented one of the core VPN protocols while working at NordVPN.
>enjoys breaking down and analysing neural networks and their architectures.
>has 50 years of coding experience.
>cofounded https://t.co/gfokj5Errj, tsar of code.

bullish on: a more algorithmic approach to AI

[condense.chat @densechat]

Meet another member of our team: @illfil_

>23 years old.
>started AI in Ukraine and moved to Lithuania.
>accidentally became a solar cell researcher for 9 months.
>missed ML too much, got into KTU's AI Academy and pivoted back for good. >walked away with two bachelor's degrees: physics and AI.
>co-authored peer-reviewed AI research while still an undergrad.
>lead scratch engineer
>cofounded https://t.co/gfokj5Errj, building AGC (agentic general compaction)

bullish on: cheaper and faster models

[condense.chat @densechat]

Its time to meet our team

Say hi to @MarioPeng242578

>26 years old.
>started researching AI in 2021 and language models before they entered the mainstream.
>studied computer science, linguistics, and psychology at UCLA.
>worked across three ML labs, researching different areas of AI and connecting ideas across disciplines.
>active in open source AI research communities.
>explored most important challenges: helping language models work effectively with growing amounts of context.
>cofounded https://t.co/gfokj5Errj, researching and building frontier systems for context compaction.

bullish on: more efficient AI

[condense.chat @densechat]

@radinoregon Claude code change some things on their side, which are now addressed and you shouldn't get errors anymore!

[condense.chat @densechat]

@hunteroor 🔥

[condense.chat @densechat]

We're giving you 100M free tokens to prove your agent is wasting context.

Most of what your agent sends upstream is dead weight. We built state-of-the-art compaction to strip it, and today we published the receipts.

One real session, run to full depth, benchmarked against headroom:
condense: 53.8% of tokens saved, 37.3% off the bill
headroom: 15.1% and 13.8%

Everything is open in the post. One curl line, and triple your Fable usage.

Full breakdown in the comments

[condense.chat @densechat]

@AkashBajwa96 We appreciate!

[condense.chat @densechat]

@ulysdev 🔥

[condense.chat @densechat]

@ukrroot It’s our own which checks impact on agent decisions and faithfulness. We will publish everything so you can rerun it and replicate yourself

[condense.chat @densechat]

@itsthedonhashim Well we didn't fine-tune LLM Lingua like majority did but instead trained our own

[condense.chat @densechat]

@radinoregon Check your DMs!

[condense.chat @densechat]

@itsthedonhashim Around 2x

[condense.chat @densechat]

@itsthedonhashim This why we train our models specifically for this task so quality impact would be minimal. And our Helena model is super fast: https://t.co/XOfFvbjh1X

[condense.chat @densechat]

@ukrroot We will😉, our open source evaluations are already in progress, and you will be able to run them by yourself

[condense.chat @densechat]

@edvard_exe @Gntrs7 Speak the truth

[condense.chat @densechat]

@geraslabaigeras 🔥

[condense.chat @densechat]

@diegocabezas01 Extend those minutes with us, we give 100M tokens free

[condense.chat @densechat]

@GalvosasJ We like to make people happy

[condense.chat @densechat]

@radinoregon We're very happy to hear that! 1M saved already is impressive, keep us posted on how it goes 😄

[Ryan DesJardins @radinoregon]

I've been testing https://t.co/PeRVYugl7r from @densechat, and so far, a saved token is a token earned, especially when running Fable. These were the tokens saved very early on in a /goal I was running, where Fable was orchestrating Opus and Sonnet. Thumbs way up so far.