Track the retirement date of every model you call

Review by

2027-06-30

How to do it

Providers retire old models, and calls to a retired model fail. Notice can be short, such as the 60 days Anthropic promises for public models. Build the list from the gateway settings, so no model is missing, including embedding, speech and fine-tune base models. Read the dates from each provider's deprecation page or API on a schedule, such as weekly, and store the date checked. Keep a short runbook per feature. It names the replacement candidates, the eval set to run and who signs off. Switching the base model is a new model, so rerun the whole eval set, not a sample, as review and test every prompt change requires. Fine-tunes are rebuilt on the new base, as rebuild fine-tuned models from stored data says. It is the model form of keep one list of support end dates.

Examples

TypeScript example

const ALERT_DAYS = 90;
const DAY_MS = 86_400_000;

export function modelsNearRetirement(models: ModelEntry[], today = new Date()): ModelEntry[] {
  return models.filter(model => model.retiresOn && (model.retiresOn.getTime() - today.getTime()) / DAY_MS < ALERT_DAYS && !model.migrationTicket);
}

Python example

ALERT_DAYS = 90

def models_near_retirement(models: list[ModelEntry], today: date | None = None) -> list[ModelEntry]:
    today = today or date.today()
    return [m for m in models if m.retires_on and (m.retires_on - today).days < ALERT_DAYS and not m.migration_ticket]

How to check

A CI test reads every model ID from the gateway settings and finds each in the list. Each entry has a retirement date or "none announced", checked in the last 14 days. A scheduled job fails when a date is under 90 days away and no migration ticket links to it.

Sources

• Claude Docs, model deprecations
• OpenAI API deprecations

Related rules

• AIF-81 Write a short card for every AI feature in AI features
• AIF-59 Pin the AI model version you depend on in AI features
• AIF-8 Review and test every prompt change in AI features
• AIF-62 Rebuild fine-tuned models from stored data and compare before switching in AI features
• ENG-71 Keep one list of support end dates in Engineering