The cache, not the router, is the line item: 84% to 94% of a routed turn's cost

The 94% is not a provider price. It is arithmetic on LiteLLM's own benchmark of its auto-router, which measured prompt caching stacked with routing across five datasets, two of which read what the provider's cache actually did ^[1]. It matters because it puts the largest inference COGS lever in the caching layer rather than in model selection.

The setup, per LiteLLM: the router arm ran one model group with four tiers, SIMPLE to claude-haiku-4-5, MEDIUM to claude-sonnet-5, COMPLEX and REASONING to claude-opus-5 ^[2]; the baseline arm sent every request to a single frontier model with prompt caching on ^[3]. The stack came in 37% to 69% cheaper than caching alone ^[4], while a router with caching switched off ran about 4x the cost of caching one fixed model ^[5].

Those two figures bracket a routed turn. A cached routed turn costs 31% to 63% of the cached single-model baseline ^[12]; an uncached one costs roughly four times that baseline, or 6.3x to 12.9x the cached turn ^[13]. Put the other way, caching removes 84% to 92% of what a routed turn costs ^[14]. Read "4x more expensive" strictly, as cost equal to five times baseline, and the top of the band moves to about 94% ^[15]. It is a band whose upper edge turns on a preposition, not a point estimate.

What it turns on operationally is that a model switch is not an eviction. LiteLLM reports that in 99.3% of 4,684 real switch-backs on its gateway traffic, the earlier model's cache was still warm ^[6], and that only 4% of cache misses were preventable by a background warmer, which it values at plus or minus two percent ^[7]. The loss comes from prefix drift instead: caching requires the start of the prompt to be byte-for-byte identical, and a reordered tool or an interpolated timestamp silently invalidates it, with cache_read_input_tokens dropping to zero as the only signal ^[8].

So the number to instrument is not routing topology. Anthropic's cache diagnostics beta compares a request against the previous response id and names the first point of divergence, whether model, system prompt, tools or message history ^[9].