Preparedness v2 + Frontier Safety v3: the dual evaluation frameworks that define how frontier labs measure capability and risk

3 sources

OpenAI Preparedness Framework v2 (risk categories, score thresholds, mitigation obligations, cross-functional review board) and Google DeepMind Frontier Safety Framework v3 (capability evaluations, early-warning indicators, mitigation deployment obligations) define the dual evaluation frameworks that frontier labs publish alongside model releases. The two frameworks differ in structure (Preparedness is risk-categorized with thresholds; Frontier Safety is capability-evaluation with early-warning) but converge on the same goal: producing a structured artifact that the lab uses to decide whether to deploy, mitigate, or pause a given capability.

Why now

Both frameworks updated in the same week — the dual evaluation framework pattern is the lens the rest of 2026 will use to read frontier lab safety communications.

Why it is worth publishing

Demo potential: side-by-side comparison of Preparedness v2 vs Frontier Safety v3 — what each framework measures, what each mitigation obligation looks like, where they converge and where they diverge.

Evidence basis

OpenAI Preparedness docs + Google DeepMind safety page + The Decoder weekly roundup

“OpenAI Preparedness v2 and Google DeepMind Frontier Safety v3 both shipped this week — and together they define how frontier labs measure capability and risk; creators should read the framework structure, not just the headline version number.”

Angle

Use the dual update to introduce the 'evaluation framework' framing — Preparedness v2 + Frontier Safety v3 together define how frontier labs measure capability and risk — and use that lens to discuss what creators should read into each framework's structure.

Format

Long-form explainer

Demo idea

Record a 10-minute explainer: 3 min on 'why dual evaluation frameworks matter' (each lab has its own structure; creators should read the structure), 3 min on Preparedness v2 (risk-categorized with thresholds), 3 min on Frontier Safety v3 (capability-evaluation with early-warning), 1 min on the convergence.

Platform notes

Specific risk categories, score thresholds, evaluation tasks, and early-warning thresholds beyond the captured summary are not extracted; do not state specific numbers or category names.

Usable claims

OpenAI updated the Preparedness Framework to v2 — risk categories, score thresholds, mitigation obligations, and cross-functional review board.
Google DeepMind released Frontier Safety Framework v3 — capability evaluations, early-warning indicators, mitigation deployment obligations.

Evidence pipeline

From the news

Breakdown

Preparedness v2 and Frontier Safety v3 are both 'evaluation frameworks' but they have different structures (Preparedness is risk-categorized with thresholds; Frontier Safety is capability-evaluation with early-warning) and different mitigation obligations. This explainer uses the dual update to show where the two frameworks converge (deploy / mitigate / pause decisions) and where they diverge (structure, naming, threshold semantics) — without claiming one is the standard the other should follow.

Sources

Risks

Docs confirm the existence of the framework versions but specific thresholds, categories, and tasks beyond the captured summary are not in this pass. Verify specific capability claim against the underlying vendor docs and the actual license / pricing matrix before stating it on the record; do not paraphrase per-platform pricing or license terms into specific dollar figures or commercial-use clauses.
Use The Decoder and IT之家 as media-type corroboration, but read the underlying lab docs for any specific safety claim before stating it on the record. Verify specific capability claim against the underlying vendor docs and the actual license / pricing matrix before stating it on the record; do not paraphrase per-platform pricing or license terms into specific dollar figures or commercial-use clauses.

Demo ideas

Side-by-side framework walkthrough: Preparedness v2 vs Frontier Safety v3 — what each measures, what each mitigation looks like.
Convergence / divergence matrix: where the two frameworks converge (deploy / mitigate / pause decisions) and where they diverge (structure, naming, threshold semantics).
Reading guide: a creator's primer on how to read each framework's structure when a frontier lab ships a model.