The Cheap Model Billed More Per Turn: Routing to Haiku Mid-Session Costs 94% More

The +94% is arithmetic on posted prices, not a benchmark. Anthropic charges $5 per million input tokens for Opus 5 and $1 per million for Haiku 4.5 ^[1], and on a long session a router that hands one easy turn down to the cheap model raises that turn's input cost by about 94% ^[11]. That is the inverse of what a difficulty router is installed to do ^[6].

The mechanism is the cache, not the model. A request whose prefix matches the previous one is billed at 10% of the base input rate ^[3], so a warm Opus prefix reads at $0.50 per million rather than $5 ^[10]. But the cache belongs to one model: Opus 5 and Haiku 4.5 have different weights and separate cache stores, and a warm prefix does not transfer ^[4]. Switching discards the prefix and forces a cold prefill of the entire conversation on the new model, so the cheap rate applies to the whole transcript rather than to the new instruction ^[5]. And every turn re-sends the full conversation, which is why a single coding task can push 400K to 2M cumulative input tokens ^[2].

Run the numbers on a turn carrying 300,000 cached tokens plus a 1,000-token instruction: $0.155 on Opus with the cache warm, $0.301 on Haiku cold, or +94% ^[11]. As the new instruction shrinks toward nothing, the penalty converges on $1 over $0.50, a hard ceiling of +100% ^[12]. Switching only pays when new tokens exceed about 11% of the turn's input ^[13], which is rare deep in a session. The five-minute default cache lifetime, refreshed free only when the entry is used ^[7], compounds it: turns sent elsewhere do not refresh the frontier prefix, so a stretch of cheap turns can also expire the expensive model's cache ^[14].

Two things move the figure: a cheap model priced under 10% of the frontier rate, which would put it below $0.50 per million here ^[13], and routers that decide at request or task boundaries instead of at each step ^[9].