A $14.34 router matched Opus-5's score on LiteLLM's 21-task benchmark

The saving is $5.40 ^[1]. Spread across 21 tasks that is about 26 cents a task ^[2], and the classifier call that produced it cost at most $0.34, which is 2.4% of the winning arm's bill ^[8][11]. Per solved task, the router paid $0.90 and Opus-5 on its own paid $1.23 ^[4]. Turning 26 cents a task into $1m of annual saving takes about 3.9 million tasks ^[3].

The context result is narrower than it sounds. All three context arms use one 4o-mini classifier and one tier map, so the only variable is how much of the conversation the classifier reads before it picks a tier ^[15]; the answering model's context window is not what moved. Reading the last three user messages instead of one raised Opus's share of requests from 34% to 45%, and LiteLLM attributes most of the extra cost to that escalation ^[9]. The bill went from $12.86 to $18.58 ^[3]. Swapping in gpt-5.4-mini on a single message bought the same 16 tasks for $14.34, or $4.24 less ^[6]. Median time per task rose from 6.8 minutes to 8.3 ^[4]. The arm that also read assistant replies sent 5% of requests to Haiku and scored worst of the three ^[6].

The arm I would look at hardest has no classifier in it. Its 87% of Opus quality at 56% of the cost ^[7] works out to roughly $11.05 and about 14 of 21 tasks ^[7][8], which is $0.79 per solved task, cheaper than the routed or the unrouted Opus figure ^[9]. The two extra tasks the gpt-5.4-mini router picks up cost $3.29 between them, $1.65 each ^[10]. LiteLLM reports cost as a per-arm total across all 21 tasks, taken from its gateway spend logs ^[14].

The scoring is soft at this size. One task out of 21 moves a solve rate by 4.8 points ^[12], and errors ran between zero and four per arm, excluded from the denominator only where a task did not complete for that arm ^[11]. Every tier is Anthropic: Haiku for SIMPLE, Sonnet for MEDIUM, Opus for COMPLEX and REASONING, with Sonnet as the fallback ^[10]. All eight arms ran the same 21 tasks with prompt caching on, which is what LiteLLM means by like for like ^[12][1], and the company published the run on its own documentation site alongside a call for design partners ^[13].

The router cannot beat Opus, because its top tier is Opus, reached on between 34% and 45% of requests ^[9]. So what is on offer is a price effect on a fixed ceiling. The claim most likely to break is the one about the single message: Terminal-Bench tasks arrive as self-contained specifications, so one user turn tells a good classifier what it is dealing with, and on a workload where the first message does not describe the job, the arm that paid $18.58 for three turns of context is the one that should win ^[3][15].