Why GPT-4.1 Costs Double Claude Sonnet Despite Cheaper Tokens

Most routing systems treat model selection as a classification problem, but the team behind this research learned the hard way that it’s really a systems optimization puzzle. Three dimensions tripped them up, starting with cost. On a benchmark of 417 agent tasks, Claude Sonnet 4.6 racked up a $79 bill — roughly half of GPT-4.1’s $155 total. That sounds insane given that GPT-4.1 has lower sticker prices on both input and output, and Sonnet needed three times as many reasoning steps. The culprit? Cache-read pricing. Agent workloads reuse massive context chunks, and Sonnet’s cheaper cached reads flipped the economics entirely. A router that only glances at a pricing sheet is effectively navigating with a broken compass.

Complexity is the second hidden trap. The intuitive play — estimate task difficulty, then route accordingly — falls apart because difficulty is often invisible upfront. A simple-sounding request like “summarize this contract” can spiral into retrieval, tool use, and multiple rounds of refinement. Meanwhile, a dense technical prompt might sail through on a small, specialized model. Even if you could perfectly predict difficulty, it’s still just one signal among many. Production routers are juggling cost, latency, compliance, data residency, and corporate approved-model lists simultaneously. The job isn’t picking a model; it’s a constant, multi-variable negotiation.

Latency also isn’t just about model size. A “faster” model can feel sluggish if the endpoint is busy, the cache is cold, or the hardware is underpowered. Routing overhead itself adds drag, especially with per-step routing that multiplies decision points. The team’s solution was to abandon the classification mindset and reframe routing as an optimization problem. Their algorithm balances cost, quality, and latency while staying lightweight — about 6 ms and 2 kB of memory per task. On the AppWorld Test Challenge, their latency-tuned configuration hit 84% accuracy for $93 and 83 seconds. That’s a 21% cost cut and 9% speedup versus running Opus alone, with only a 4-point accuracy sacrifice. A standard difficulty-based router landed in a similar accuracy range but at a higher cost, never fully exploring the tradeoff frontier. The fundamental lesson: routing is about optimizing systems, not picking winners.