Breaking

condense.chat opens compression proxy that shrinks input tokens for coding AIs

July 3, 2026 at 09:09 EDT

AI Agents
Software Dev & Coding
Infra & Chips

Around July 3, 2026, condense.chat (@densechat) opened general access to a "compression proxy" that sits between a coding AI agent and the LLM, compressing input tokens to cut API bills. It works as a drop-in that only requires changing an SDK's base_url, and the company claims bill reductions of over 72% in real sessions.

condense.chat · Compression Proxy

Squeezing the context: a proxy that cuts AI token costs by ~70%

A drop-in proxy sits between a coding agent and the upstream LLM, compressing prompts, retrieved documents and tool outputs — claiming big cuts in context and cost while keeping accuracy intact.

−64%

Context reduction

−70%

Token & cost reduction

19.3%

Faster responses

Tokens reaching the LLM

Same task, same unit — the proxy removes roughly 70% of tokens before they hit the model.

100%

Uncompressed

≈30%

Via proxy

CoQA · 150 turns

92.0%

accuracy in auto mode — above the 90.0% uncompressed baseline, while cutting input tokens 30.6%.

LongBench v2 · up to ~500k

49.3%

accuracy with a 38.8% token cut — and 44.0% vs 40.0% on code & structured-data.

How it fits in — just change the base_url

Coding agent
Claude Code · Codex · OpenCode

→

Compression proxy
Helene 1 · Adeline 1 · zero-retention

→

Upstream LLM
tens-of-ms latency added

Clear upside

Everyday Claude Code users report real-world token reductions above 50%, with accuracy held or improved and free tokens offered in the early phase.

Still early

Long-term usage data is limited as general availability has only just begun; access needs an API key, possibly under an approval process.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.