We've updated our Responsible Scaling Policy (RSP) to reflect new capabilities in interpretability and eval design for frontier models.

What changed

The core structure remains the same: we commit to specific safety and reliability milestones before scaling compute. What changed in this version:

Tighter interpretability requirements. Previous RSP versions required "reasonable efforts" to understand model internals. v2 requires demonstrated mechanistic interpretability progress — specific circuits identified, specific behaviors localized — before crossing certain capability thresholds.

Expanded eval coverage. We now require multi-domain evals covering not just standard capability benchmarks but: reasoning under distribution shift, robustness under adversarial inputs, and novel "catastrophic risk" evals that probe for emergent dangerous capabilities.

Third-party auditing. Starting next quarter, our RSP commitments will be audited by an independent safety evaluator. Results will be published publicly.

Why interpretability matters for scaling decisions

The honest answer: we don't fully understand why large language models work as well as they do. But we're getting better at understanding when they'll fail in ways that matter. Before scaling to the next order of magnitude of compute, we want to have mechanistic insight into the failure modes — not just statistical assurance from benchmark performance.

This is harder than it sounds. Mechanistic interpretability research is in its infancy. But the alternative — scaling blindly based on benchmark performance — is riskier.

What's next

We'll publish quarterly RSP progress reports starting next month. The next major RSP update is planned for Q4 2026, coinciding with our expected next major training run.

You can read the full updated policy at our safety page.