How Parallax cut their Claude bill 47% without a single customer complaint
The routing rules, prompt caching, and quality guardrails that took $1.2M/yr in Anthropic spend down to $640k — with faster responses.
When our weekly Anthropic invoice crossed $24,000 in March, the CFO asked a fair question: do we actually need Sonnet for every request? The honest answer was no — but we had no instrumentation to prove it, and no safe way to roll back if quality regressed.
We dropped SpendTensor in front of our gateway in a single afternoon. Within 48 hours we had token-accurate traces on every call: provider, model, prompt fingerprint, cache hit rate, p99 latency, and downstream eval score. The picture was uncomfortable. 71% of our Sonnet traffic was classification and summarization — workloads that Haiku handled at 4% of the cost with no measurable quality drop.
The win wasn't a single model swap. It was the routing layer: prompts get fingerprinted, scored against a small eval set, and routed to the cheapest model that clears the quality bar. If the cheaper model regresses on a live request, we fall back to Sonnet in <5ms. The customer never sees it.
Three months later, we've cut spend 47%, sped up p50 latency by 380ms, and shipped two new features inside the budget headroom. The board stopped asking about the AI bill.