Long-horizon eval
+38.4%
Task completion on HorizonBench-500 over Halyard 2, with no change to latency at p50.
A 1.2M-token context, verified tool use and a planning head trained on 41 billion steps of agentic traces. Available today in the API and on-prem.
+38.4%
Task completion on HorizonBench-500 over Halyard 2, with no change to latency at p50.
Dr. Ines Achterberg, Head of Reasoning
Plan a 3-week migration of our billing service to event sourcing, with a rollback at every step.