Guardian 2.0 Thinking

Today we are releasing Guardian 2.0 Thinking — the second generation of our flagship reasoning model, and a substantial upgrade across every dimension that matters for production AI workloads.

A standout strength is instruction-following: Guardian 2.0 Thinking scores 76.5 on IFBench — the benchmark that most directly measures whether a model does exactly what you ask, with nothing missed and nothing invented. For production deployments where reliability matters as much as capability, this is the metric that counts.

Guardian 2.0 Thinking scores 86 on GPQA Diamond (graduate-level science reasoning), 85 on MMLU-Pro (professional knowledge across law, medicine, engineering), and 93 on AIME 2026 (competition-level mathematics). It natively supports 200+ languages — exceptionally broad multilingual coverage.

The three-level reasoning system carries forward from 1.0 with improved calibration. Low-reasoning mode is substantially faster. High-reasoning mode shows the largest gains on multi-step scientific and mathematical tasks. Developers control the speed-accuracy trade-off per request.

All inference follows Guardian's privacy-first architecture — encrypted in transit and at rest, with minimal retention by default. Nothing has changed about our privacy architecture — we have only made the model better.

Guardian 2.0 Thinking is available now through the Onora App. Existing users have been automatically upgraded. New users can get started at onora.app.