OpenAI Path to Astra document lists 12 capabilities and 8 safeguards for frontier model release

The document enumerates capability thresholds drawn from OpenAI's 2023 Preparedness Framework and internal evaluations of o1-preview runs. It requires evidence of containment for each listed risk before external deployment or expanded API access. No quantitative benchmark scores or pass/fail criteria appear in the released text.

Cross-reference with Anthropic's Responsible Scaling Policy v1.0 shows overlap on replication and deception vectors but diverges on third-party audit mandates. OpenAI omits the external review board structure present in the Anthropic document and provides no schedule for third-party verification of its internal red-team results.

Operational effect is restriction of model weights and fine-tuning endpoints to vetted partners until the listed safeguards are attested. This aligns with prior pattern of incremental API tiering rather than full open release seen in earlier GPT series launches.

Next milestone is the scheduled internal audit cycle referenced in section 4.3, expected to produce an updated capability scorecard within six months of any new model training run exceeding current o1 metrics.