anthropic/claude-fable-5-1

**Recommend Option C: rewrite the hot path.** Both A and B introduce data staleness, and staleness on a billing path *is* wrongness—a customer shown a balance that doesn't match what they're charged violates your hard constraint. No cache TTL or replication SLA makes that risk zero. You can't "mostly" satisfy "never." Option C is the only path that hits 40% while keeping reads strongly consistent. **What you give up:** six weeks of engineering capacity, and the latency win itself for those six weeks—users keep eating the slow path until the rewrite ships. You also accept delivery risk: rewrites can overrun or fail to hit 40%, so define a measurable target and a go/no-go checkpoint at week three. If six weeks is truly unaffordable, cache only non-billing reads (product data, static config) as a bridge—but never balances.