AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Mid-Task Model Switches Carry a Hidden “Handoff Tax,” Study Finds

Coding agents often run for dozens of model calls, and users routinely switch models in the middle of a task: escalating from a cheap model to a stronger one when it struggles, or downshifting once the hard reasoning is done. A new study from researchers at AWS quantifies a hidden cost of those switches, which they call the “handoff tax.” When a model is asked to continue a trajectory produced by another model, the inherited context can burden rather than help the receiver.

The researchers ran 58,000 agent episodes on SWE-bench Verified, a benchmark of real GitHub issues, using pairs of models from the Claude and GPT families. In each family, LC (low-cost, low-capability) and HC (high-cost, high-capability) denote the cheaper and stronger models: Haiku 4.5 and Opus 4.7 for Claude; GPT-5.6 Luna and Sol for GPT. They varied switch direction, timing, and interface. All interfaces preserved the edited working tree on disk; they differed only in what conversation history the receiver inherited: the full trajectory (Raw), a summary written by either model (Compactpre or Compactsuf), or no trajectory at all (Traj-drop).

The central result is blunt. For Raw escalation, the default continuation interface considered by the study, the HC receiver recovered less than half of the gap between LC-only and HC-only pass rates: 47% for Claude and 36% for GPT. The clearest failure appeared in Claude escalation. Raw handoff solved 69.2% of tasks at $1.61 per run, versus 79.2% and $0.72 for Opus from the start—a 10-percentage-point gap in pass rate. Even after paying for the Haiku prefix and throwing away its edits and conversation, the “Abort + HC fresh” control cost $0.90 and solved 79.2%—strictly better than Raw on both cost and accuracy.

The value of inherited context reversed with direction. Downshift also offered a more favorable cost-quality point: for Claude, Raw handoff from Opus to Haiku raised pass rate from 54.6% to 65.6% while retaining 80% of LC’s cost advantage; for GPT, it retained 79% of HC’s quality advantage. The interface reversal was sharper. In GPT escalation, dropping the LC trajectory while keeping its code edits raised quality recovery from 36% to 84%, though it remained slightly costlier than running Sol alone. In downshift, removing the HC trajectory hurt the LC receiver: for Claude, Traj-drop recovered only 28% of HC’s quality advantage versus 50% for Raw. The authors explain this as a directional duality: LC trajectories burden HC receivers, whereas HC trajectories guide LC receivers.

The study has important boundaries. It tests only two model families and one coding benchmark; the hard-task subset, where reduced-context escalation looked attractive, was small (roughly 24 tasks per cell) and exploratory. Runs were single episodes, so variability across repeated attempts was not estimated, and dollar conclusions depend on provider pricing and cache rates.

Still, the findings suggest handoff design deserves the same attention as model selection. Routing decides which model acts next; the interface decides what that model inherits. Treating the full trajectory as a default, the authors argue, is not a neutral choice—and in long-running agents, how a handoff is packaged can matter as much as who receives it.