Every serious AI tool now has a model switcher. You start on the cheap, fast model, it gets tangled halfway through a task, and you flip to the expensive one to rescue it. The expensive model then inherits everything: the half-finished work, and the whole conversation that produced it, including the cheap model's guesses and dead ends. A team at AWS measured what that inheritance costs. Their headline is the one worth repeating to anyone who pays an AI bill: in one of the two model families they tested, abandoning the cheap run and restarting the strong model from scratch came out both cheaper and better than letting the strong model carry on, even after paying for the wasted work.
The handoff tax, named.
The paper calls it the handoff tax: the quality you fail to recover, and the money you overspend, when one model has to continue a run that another model started. The study is large for this kind of question: 58,000 agent runs and 36 billion tokens, with the switch forced at seven different points in a task, in both directions, through four different ways of passing the baton. The authors draw one distinction worth keeping: model routing decides which model acts next, and the handoff decides what that model inherits.
Those four ways are the part an operator controls, so they are worth naming plainly. Every one of them kept the work already saved to disk. What changed was how much of the first model's thinking came along.
- Full transcript. The receiving model sees everything the first one said and did. This is what you get when you flip the model inside the same conversation.
- Summary by the outgoing model. The first model compresses its own run into a summary, and only the summary crosses.
- Summary by the incoming model. The receiver reads the run, summarizes it, and continues from the summary alone.
- Files only. The conversation is dropped. The receiver starts from the saved work and the original task.
For escalation, the authors added a control: abort, throw away the cheap model's work entirely, and run the strong model from the original task, while still charging for everything the cheap model had already burned.
What each handoff recovered, going up.
Quality here is measured against the gap between the two models. If the cheap model alone solves some share of the tasks and the strong model alone solves more, 0% means the handoff landed at the cheap model's result and 100% means it reached the strong model's. Cost is shown against the obvious alternative: running the strong model from the start.
Read the top row first, because it is the default. Flip the model mid-conversation and the strong model recovered 47% of the gap in the Claude pair (36% in the GPT pair), while the Claude run cost more than twice what starting strong would have. The bottom row throws the cheap work away, pays for it anyway, and still reaches full strong-model quality for about a quarter more than starting strong. On those numbers, carrying on with the transcript was worse than starting over on both counts. The authors' word for it is dominated.
The middle rows are the fix when the work is worth keeping. A summary written by the outgoing model was the cheapest way up, close to the cost of starting strong. Dropping the conversation and keeping only the files recovered the most quality of any continuation. The extra cost comes from rereading: carrying the full transcript made each of the strong model's later steps about 2.2 times as expensive as carrying the outgoing model's summary in the Claude pair (1.6 times in the GPT pair), for roughly the same number of steps. You pay the strong model to reread the cheap model's chatter on every turn.
Why direction flips the answer.
This is the paper's most useful result. Going down, from strong to cheap, the trail that hurt on the way up becomes an asset. When the strong model did the first part of a task and a cheap model finished it, keeping the strong model's trail mattered. In the Claude pair, the cheap finisher kept 50% of the quality gap with the full trail and 28% with files only. In the GPT pair, 79% with it and 53% without. Without the notes, the cheap model also took about 1.6 times as many steps as with the strong model's summary in the Claude pair (2.0 times in the GPT pair), rebuilding reasoning the notes would have handed it.
The office version is familiar. Hand a senior consultant the junior analyst's full meeting transcript and you have bought an expensive afternoon of reading someone else's wrong turns. Hand them a one-page brief and they get to work. Hand the junior the senior's working notes and they finish the job in the senior's direction. Take the notes away and they rebuild the reasoning from scratch, slower and worse. Context written by the stronger model is an asset to a weaker one. Context written by the weaker model is a weight the stronger one pays to carry.
Two rules for any tool with a model switcher.
The study measured coding agents. The rules below are our translation to the other places the same switch happens: a chat assistant drafting a proposal, a research assistant halfway through a report, an automation that falls back to a bigger model on a hard case. The mechanism, one model rereading another model's conversation, is the same. The paper's side tests outside coding are a caution on that step: when the conversation itself held the requirements or the research findings, the transcript was worth far more (see the limits). So treat these as defaults to test on your own work, not as laws, and put anything the conversation learned into the brief.
- Escalating: start the stronger model clean, with a brief.
Do not flip the model inside the same conversation. Open a fresh session on the stronger model, attach the work so far, and write a few lines: the goal, what is done, what was tried and failed, and what you need next. If the brief is hard to write, or the work so far looks shaky, restart the strong model on the original task instead. In one family that beat every continuation on cost and quality at once.
do · carry the work and a brief up, leave the transcript behind - Downshifting: hand the notes down with the work.
When the strong model has done the hard thinking and you move the rest to a cheaper model to save money, keep the trail. Either switch inside the same conversation, or have the strong model write its plan and reasoning into a note first and give that note to the cheaper one. A cheap model handed only the files spent more steps and kept less of the quality.
do · carry the work and the strong model's notes down
One inference sits on top of the two rules, and it is ours rather than something the paper tested head to head. In the Claude pair, every way of escalating mid-run cost more on average than simply starting strong, while handing down from a strong start gave a favorable middle point in both families. So when you can tell in advance that a task is hard, starting on the strong model and handing down is the safer bet on quality than starting cheap and paying for a rescue. The one hint the other way, on cost, is exploratory and sits in the limits below.
Where this hit our own routing.
We run this pattern every day in our own tooling: cheap tiers for mechanical steps, stronger tiers for judgment, and an escalation step when a cheap run struggles. Our written default assumed that escalating mid-run was cheap. This paper measures that assumption and partly refutes it. The half that holds up is how we already handled downshifts: plans and notes written by the stronger tier travel with the work to the cheaper one. The half that needs rewriting is escalation. Completed work should stand, the conversation should stay behind, and the stronger model should start from a brief.
It also sharpens two older notes. Per-step spend, the instrument from making cost a first-class metric, is exactly where the handoff tax shows up: the same number of steps, each one dearer. And the finding that stale context competes with the current goal has a cousin here. The competing context is another model's reasoning, and it competes hardest when the reader is the stronger model.
What this does not prove.
The switches were scheduled, not earned. The authors forced each switch at a fixed point in the run, not when the cheap model was actually struggling, which is how people really escalate (and spotting a genuinely stuck run is its own problem, see the stalled-agent note). A run that is visibly stuck may carry worse baggage, or more useful failure information. The study cannot say which.
The cost verdict belongs to a cost-per-task ratio. The two Claude models sat roughly 2× apart in cost per task when measured, the two GPT models roughly 8×. That is why the full-transcript escalation was a clear loss in one family and merely mediocre in the other. Prices move. Rerun the arithmetic with your own, as we argued when the open tier got cheaper.
The hard-task exception is a hint. On the hardest tasks in the Claude pair, the slimmed-down escalations became cheaper than starting strong while recovering 65–74% of the gap, and the full-transcript handoff never did. The authors call this exploratory, with about 24 tasks per cell, and so do we.
Outside coding, the transcript earned its keep. The authors ran two side tests with the full transcript only. In a chat setting where requirements arrive turn by turn (Claude pair), escalating recovered 86% of the quality gap and still cost less than starting strong, while handing down kept only 31%; the authors' explanation is that the task only became solvable once the last requirement arrived, and the strong model was there to receive it. In a web-research setting (GPT pair), escalating recovered 96% of the gap but cost more than aborting and restarting the strong model. Neither test tried a brief or files only, so neither says the brief loses. They do say a cheap model's conversation is not always dead weight: when it holds facts the saved work does not, those facts belong in the brief.
If your team routes work between a cheap model and a strong one and the bill keeps drifting, the contact form is the fastest way in. Tell us where the switches happen, and we will send back a written read on which handoffs are leaking cost, free.