What they did
The authors paired lower-cost, lower-capability and higher-cost, higher-capability models from the Claude and GPT families. During multi-step coding tasks, they varied when models were switched and whether the switch escalated to the stronger model or downshifted to the cheaper one.
They also varied the handoff interface: transferring the complete interaction trajectory, compacting it, or removing it while retaining the resulting repository state. This separated the value of prior conversational and tool-use history from the value of the code changes already made.
Key findings
- Full-trajectory escalation recovered less than half of the quality gap between lower- and higher-capability models, despite a substantial cost premium.
- Reducing the weaker model’s inherited trajectory improved escalation performance compared with passing it in full.
- Downshifting after stronger-model work produced a more favorable cost-quality point than escalating after weaker-model work.
- The preferred handoff interface depended on direction: removing stronger-model trajectory information harmed downshift performance, reversing the pattern observed for escalation.
Why it matters
The results challenge the assumption that model routing can freely combine cheap execution with strong-model reasoning. Agent harnesses may need direction-specific handoff policies: stronger models can benefit from reconstructing the task from repository state rather than inheriting a weaker model’s reasoning, whereas cheaper models may need the richer trajectory left by a stronger predecessor.
Caveats
The abstract does not report the coding benchmarks, exact model versions, switch points, absolute success rates, or cost figures, limiting conclusions about effect size and generalizability. The study covers Claude and GPT model pairs in coding-agent settings; whether the same handoff tax appears in other agent tasks, model families, or substantially longer runs remains open.