DeepMind Frontier Safety Framework v3 — Gemini 3.1 Ultra deployment review references public capability thresholds, completing three-lab procedural convergence
DeepMind released Frontier Safety Framework v3 on May 28, with the Gemini 3.1 Ultra deployment review the first application of the updated procedure. The framework adopts publicly-stated capability thresholds that mirror Anthropic's Mythos restricted-release methodology and OpenAI's Superalignment Report, completing the three-lab procedural convergence on capability-driven release gating as the deployment-control default.
The framework update is the substantive piece. DeepMind's prior Frontier Safety Framework iterations (v1 in 2024, v2 in 2025) covered cyber, bio, and persuasion capability axes but used internal-only thresholds for the release-gating decision. v3 publishes the thresholds explicitly, in the same format as OpenAI's Superalignment Report and Anthropic's Mythos methodology. The Gemini 3.1 Ultra deployment review is the first model release evaluated under the v3 procedure — the review document is published alongside the framework and shows the capability-evaluation findings on each axis, the threshold comparison, and the deployment-control decision (deploy publicly, with deployment-policy controls on a small number of cyber sub-tasks).
The three-lab convergence is what makes the moment structurally meaningful. OpenAI's Superalignment Report and Anthropic's Mythos methodology are the other two pillars; DeepMind's Frontier Safety Framework v3 is the third. For the first time, the three major Western frontier labs use procedurally-aligned capability-thresholds-and-release-gating frameworks. The EU AI Act GPAI Code of Practice final form, due May 28, will reference the converged procedural baseline. The regulatory-procedural gap — labs voluntarily publishing thresholds versus regulators specifying them — has narrowed dramatically over the month.
DeepMind — Frontier Safety Framework v3 release May 28 2026 → · Google — Gemini 3.1 Ultra deployment review under FSF v3 → · arXiv — Frontier safety framework procedural convergence 2026 →