H1 2026 safety report
What we found red-teaming Cairn-3 for six months, and the mitigations that made it into the release.
What's new?
Subscribe to running log of model releases, product updates, and research notes from the team building Cairn.
Latest updates
Prompt caching is now on by default
Repeated context is cached automatically across requests — up to 90% cheaper and 4x faster for long prompts.
Voice mode enters beta
Talk to Cairn with sub-second latency and interruptions that feel natural, not scripted.
Our internal eval suite is now open source
The harness we use to measure reasoning, honesty, and tool use — free for anyone to run against any model.
Cairn now reviews pull requests
Connect a repo and get inline review comments that understand your codebase, not just the diff.
Structured outputs are generally available
Guarantee JSON that matches your schema on every response — no more retry loops around malformed output.
Batch API drops to half price
Queue non-urgent workloads and get results within 24 hours at 50% off standard rates.
A clearer look at how Cairn reasons
New interpretability tooling lets you trace which inputs shaped a given output, down to the token.
Context windows now hold a full codebase
Up to 1M tokens of context, with retrieval quality that holds steady all the way to the edge.
Cairn-3 ships with native tool use
Our latest model calls tools mid-reasoning instead of waiting for a final answer — faster agents, fewer dropped steps.
- Batch API drops to half price
Queue non-urgent workloads and get results within 24 hours at 50% off standard rates.
- A clearer look at how Cairn reasons
New interpretability tooling lets you trace which inputs shaped a given output, down to the token.
- Context windows now hold a full codebase
Up to 1M tokens of context, with retrieval quality that holds steady all the way to the edge.
- Cairn-3 ships with native tool use
Our latest model calls tools mid-reasoning instead of waiting for a final answer — faster agents, fewer dropped steps.
- H1 2026 safety report
What we found red-teaming Cairn-3 for six months, and the mitigations that made it into the release.
- Prompt caching is now on by default
Repeated context is cached automatically across requests — up to 90% cheaper and 4x faster for long prompts.
- Voice mode enters beta
Talk to Cairn with sub-second latency and interruptions that feel natural, not scripted.
- Our internal eval suite is now open source
The harness we use to measure reasoning, honesty, and tool use — free for anyone to run against any model.