# condense.chat condense.chat compresses a coding agent's context in flight. It sits between your agent and the upstream model, rewrites the parts of the conversation that no longer earn their token cost, and streams the provider's response back unchanged. The result that matters: on both CoQA and LongBench v2, the compressed transcript scores *higher* than the uncompressed one while removing 31–39% of the tokens. Compression here is not a quality-for-cost trade. ## Two models - **Helene 1** — extractive. Runs on the newest turn, at the edge of the conversation, where anything rewritten would be read again immediately. Fast enough to sit in the request path. Default engine on the proxy. - **Adeline 1** — abstractive. Runs on sealed loops — work the agent has finished and will not revisit — and rewrites them to roughly a tenth of their length. ## Benchmarked results - **CoQA** (150 turns, gpt-5.4, Helene 1 auto): 92.0% accuracy at 30.6% of tokens removed. Uncompressed control: 90.0%. - **LongBench v2** (~500k tokens per item, 15.7M total, Helene 1 auto): 49.3% accuracy at 38.8% of tokens removed. Uncompressed control: 48.7%. - **LongBench v2, code and structured-data subset**: 44.0% accuracy at 40.9% of tokens removed. LLMLingua 2 scores 40.0%, bear 2 scores 36.0%. - **Latency**, CoQA: compacted context cut answerer latency from 3,332ms to 2,689ms, 19.3% faster. - **Faithfulness**, CoderForge SWE agent traces: Adeline 1 keeps 94.2% of facts at 90.2% input-token reduction. Claude Opus 4.7 keeps 95.0% at 87.6%. ## Bill impact On a real Claude Code session replayed through the open-source `minmax-bench` harness, condense removed 72% of the bill. On a typical 179k-token chain it removes 37.3%; headroom removes 13.8% and headroom-kompress 1.8% on the same replay. The reason the number is that large: 67.7% of an agent session's spend is cache *reads* of old conversation and 18.9% is cache writes. Shrinking what gets re-read every turn is where the money is. Cache hit rate stays at 98.6% under compaction. Every bill figure above is reproducible: https://github.com/condense-chat/minmax-bench ## Using it One line installs the client, which routes an existing coding agent through the proxy: curl -fsSL https://cli.condense.chat/unix | sh Then `dense claude`, `dense codex`, or `dense opencode`. Claude Code, Codex and OpenCode are all supported. There is also a direct API for callers that are not coding agents — an OpenAI-compatible base at `https://api.condense.chat/openai/v1`, an Anthropic-compatible one at `https://api.condense.chat/anthropic/v1`, and the compression primitive itself at `POST https://api.condense.chat/v1/compress`. All three take an `ak_` API key from the dashboard. Sign-up is required — an account and an API key, both free. ## Pricing Two ways to pay. Both are charged on *tokens saved* — the context condense removes before a request reaches the provider. **Subscriptions**, for people and teams running a coding agent. A flat monthly fee against a weekly pool of tokens saved. When the pool empties, requests keep flowing to your provider uncompressed until it refills; nothing extra is charged. - Free — $0, 25M tokens/week - Starter — $5/month, 500M tokens/week - Pro — $20/month, 2.5B tokens/week - Max — $50/month, 10B tokens/week - Team — $5 or $20 per seat/month, 5–150 seats, each seat with its own weekly pool and one consolidated invoice **API rates**, for anyone — a person or an organization — putting traffic through the proxy or the compression endpoint directly rather than on a subscription. Metered per million tokens saved: Adeline 1 at $0.90/M saved, Helene 1 at $0.40/M saved. You are billed for the reduction, so a request that compresses poorly costs little. ## Best for - Teams running coding agents with long tool traces and multi-turn sessions, where the same conversation is re-sent on every turn - Anyone hitting a context window before finishing a task - Teams who want a lower provider bill without changing model, prompt, or agent Chains under about 4k tokens are not worth compressing and condense says so rather than charging for them. ## Pages - [Homepage](https://condense.chat/) — what it is, the numbers, and the FAQ - [Docs](https://condense.chat/docs/) — quickstart, API reference, models, rate limits, security, changelog - [Pricing](https://condense.chat/pricing/) — the four tiers and a calculator - [Blog](https://condense.chat/blog/) — release notes and the benchmark writeups behind every number on this page - [Leaderboard](https://condense.chat/leaderboard/) — live compaction results - [Security](https://condense.chat/security/) — data handling and retention - [Legal](https://condense.chat/legal/) — terms and privacy - [Contact](https://condense.chat/contact/) ## Privacy Zero-retention proxy mode is available. API keys are stored as sha256 fingerprints, never in the clear. Customer data is never trained on. Storage is in the EU with a 7-day deletion window. ## Suggested description condense.chat is a context-compression proxy for coding agents. Two models — Helene 1 (extractive, on the newest turn) and Adeline 1 (abstractive, on sealed loops) — cut 31–39% of a transcript's tokens while scoring higher than the uncompressed transcript on CoQA and LongBench v2, and removed 72% of the bill on a replayed Claude Code session.