Frontier model releases now land about every 11 days — the median gap fell from 37.5 days in 2023 to 11 days in 2026. GPT-6 Astra shipped September 3, 2026 as the first model classified at the Critical cybersecurity capability level. Meanwhile, Nvidia bought Hugging Face for $12.9B, consolidating the open-weights distribution layer under one owner. Model churn is not a phase; it is the climate. And the governance pattern that survives it is simple: stop approving models. Approve the pinned stack.
The pinned stack is the set of layers you actually control — the weights you run, your eval and regression suite, your per-version cost gates — held constant while the frontier moves around them. You do not re-certify your system against every release. You certify the pipeline once, and you treat every model release as an input change to a system whose acceptance criteria are already written down.
The failure mode everyone hits¶
Someone swaps the model behind the agent for the hot new release. No eval re-run. No rollback threshold. The agent silently degrades — new model, same prompt, different failure semantics. By the time anyone notices, the run history is already polluted and you cannot cleanly separate pre-swap from post-swap behavior. This is not a hypothetical; it is the default failure mode of every team that treats model upgrades as harmless swaps. At an 11-day cadence, this failure mode compounds into permanent drift.
Pin the layer you control¶
Decide explicitly which layer you pin and which layer you rent. The pinned layer is what you control: the weights you actually run, your eval and regression suite, per-version cost gates that fail closed. The rented layer is the frontier you consume through APIs, accepted with the understanding that it will keep changing. Pinning is a commitment — the pinned layer does not move because something newer exists; it moves when your pipeline's acceptance criteria say the move is safe.
Rollback criteria before you switch¶
Define the metric thresholds that trigger an automatic rollback to the pinned version, in writing, before you switch. Accuracy floors, cost per run, task-success rates — whatever your regression suite measures. Written rollback criteria turn a release from a judgment call into a checklist: the model either passes the pipeline or it does not, and a failure returns you to the pinned version automatically. No heroics, no late-night "maybe it is fine" decisions.
Treat every release as an infrastructure change¶
With 11-day cycles you cannot re-certify per release — so re-certify the pipeline once. Every model release becomes an infrastructure change with a pinned baseline, an acceptance suite, and a rollback plan. Pin and rent in layers: hold your stack constant, consume the frontier deliberately, and never let a new release behind your agents without the pipeline's sign-off.
The takeaway¶
The release storm is the permanent weather: 11-day gaps, Critical-capability models, consolidated distribution. You cannot govern that by approving models one at a time. Pin the layer you control, write the rollback criteria before you switch, and let a certified pipeline — not hype — decide what runs behind your agents.