Changelog

Notable changes to the proxy, the models, and the platform, newest first. Upcoming removals are on the deprecations page; long-form write-ups live on the blog.

July, 2026

Update

Developer docs redesign

Rebuilt the docs around a proper API reference: two-pane endpoint sections with request and response side by side, collapsible body parameters, and a python / javascript / go / rust / curl selector on every code sample. Added dedicated Rate limits, Security, Changelog, and Deprecations pages.

June, 2026

Release
helene-1

Helene 1 generally available

The fast, accuracy-first compaction model helene-1 went GA and became the proxy default, so existing sessions started using it automatically. On a standard question-answering benchmark, a model answering from its compacted context scored higher than the same model reading the full transcript, while sending far fewer tokens. An optional compression_rate trades savings against fidelity. Numbers in the announcement.

Update

Dashboard analytics behind a first-party relay

Product analytics for the dashboard now run through a first-party /relay proxy, so measuring usage adds no third-party requests from your browser.

Feature

Organizations

Team workspaces: a shared credit pool, per-member budgets, and an org-wide spend cap you set. The dense CLI installs once per machine and members authenticate into the org. See Built for teams.

May, 2026

News

AWS Global Startup Program

condense joined the AWS Global Startup Program, backing the GPU capacity behind the Adeline compaction family. Over one hackathon weekend, 30 teams pushed roughly 4 billion tokens through condense. Context in the note.

Release
adeline-1

Adeline 1

The heavy-lift compaction model adeline-1 shipped for long agent traces. It resolves the compacted rewrite in a handful of parallel passes instead of token by token, which is where its latency win comes from. See Adeline 1 ships.

Feature
/anthropic/v1/messages /openai/v1/chat/completions /openai/v1/responses

Provider routes

Point an Anthropic or OpenAI SDK base URL at condense and everything else stays the same, Codex and Claude Code included. Streaming and tool calls round-trip; the Responses route carries Codex. Shipped alongside Adeline 1.

Release
dense

Hello, condense

First public release of the condense proxy and the dense CLI, one login that routes Claude Code, Codex, and OpenCode through condense. Read the launch note.

Feature
/v1/compress

Direct compression endpoint

POST /v1/compress takes a transcript and hands it back compacted in the same shape, with no upstream model call, N messages in, N messages out.