LLM Orchestration · Cost Analysis

The comparison
nobody ran until now.

First surfaced in Tandemly Briefing — 2026-08-01.

Every published evaluation of multi-call orchestration compared a carefully tuned orchestration pipeline against a plain, unoptimized single-call baseline. Leins, Pelleriti, Gonnermann-Müller and Pokutta changed that. When every method, including the simplest single-call approaches, gets optimized with equal effort, the accuracy advantage of orchestration shrinks to 4.6 percentage points, at roughly two to four times the token cost.

Core finding
Orchestration does gain accuracy when compared fairly, but the gap is small and the cost multiplier is real. No evidence supports the common assumption that harder tasks will widen that gap enough to justify the overhead.
scroll to explore

A literature built on
unequal comparisons.

The case for multi-call orchestration has been built on comparisons that gave one side an advantage the other side never had: tuning.

Multi-call LLM orchestration covers a family of techniques that run the model more than once to improve a final answer. Self-Refine asks the model to critique and revise its own output. Best-of-N runs the model many times and selects the strongest answer. Debate pits multiple model instances against each other. All three have accumulated positive evidence in the published literature.

There is a problem with that literature. The orchestrated methods are usually prompt-tuned, temperature-tuned, and configuration-tuned before the comparison. The single-call baseline is often presented as-is, with the default system prompt and no optimization effort invested in it. The reported accuracy gap includes the gap from unequal tuning, not just the gap from the architectural difference.

This is not a minor methodological point. A well-tuned single-call prompt can behave quite differently from a default one. If a team decides to add an orchestration layer because the published literature shows it outperforming a single call, they may be paying for a gain they could have achieved with better prompting alone.

The question this paper asks

What happens to the orchestration advantage when every method, including the single-call baselines, is optimized to the same standard under the same budget? Is the gain from architecture, or is it partly from the asymmetry in how the two sides were prepared?

Five backbones, three domains,
equal optimization budgets.

The design is straightforward: take the same evaluation protocol, apply the same optimization effort to every method, and compare the results. The difficulty-stratification is the piece that makes the hardness question answerable.

The study compared three orchestration families against two single-call baselines. The orchestration methods were Self-Refine (the model critiques and revises its own answer), Best-of-N (the model answers multiple times and the best is selected), and Debate (multiple model instances argue toward a consensus). The baselines were task-only (a direct prompt with no chain-of-thought) and chain-of-thought (the model is asked to reason step-by-step before answering).

Every one of these five configurations was optimized using GEPA, a gradient-free procedure that adjusts prompts and hyperparameters, under an identical evaluation budget. Neither side was given more optimization effort than the other. This is the key departure from standard practice.

The evaluation covered five backbone models and three task domains: competitive programming problems, chess puzzles, and mathematics. The test items were stratified by difficulty so the authors could ask specifically whether harder problems show a larger orchestration benefit, a hypothesis that often serves as justification for the added cost.

Standard practice
Tuned orchestration vs. plain baseline. The orchestrated arm is configured and prompt-tuned. The single-call baseline is a default prompt. Any reported advantage includes the benefit from the unequal tuning.
This study's design
Tuned orchestration vs. tuned baseline. GEPA optimization is applied to all five methods under a fixed budget. Both sides are as good as the optimization procedure can make them. The comparison is then between architectures, not preparation.
What difficulty stratification adds

Breaking the test set into difficulty bands allows a direct test of the hardness hypothesis: that orchestration earns its cost on the hard tail where single-call reasoning struggles most. If the accuracy gap from orchestration grows with difficulty, that hypothesis holds. If it does not, the common justification for adding orchestration on hard tasks is unsupported.

Small gains. Real costs.
No hardness payoff.

When the comparison is fair, orchestration still wins. But the margin is small, the token cost is substantial, and the two assumptions most commonly used to justify orchestration both fail.

Max accuracy gain over optimized CoT
4.6pp
The largest average accuracy improvement from orchestration over an equally optimized chain-of-thought baseline, across all methods, domains, and backbones.
Max accuracy gain over optimized task-only
4.5pp
Gain over the simplest equally-optimized baseline: a direct task prompt with no reasoning step. The gap is comparable to the CoT comparison.
Token cost multiplier
2–4x
The mean total token cost of orchestrated methods relative to single-call baselines. This is not inference cost at the architecture level. It is the measured cost in actual tokens spent.
Hardness interaction
None
Within-benchmark analysis found no evidence that orchestration methods gain more accuracy over single-call baselines on harder items. The hardness-justification assumption is not supported.
Finding: Method choice does not transfer across backbones

Mixed-effects analysis found strong method-by-backbone interaction in all three benchmark domains. An orchestration strategy that outperforms the baseline on one model gives no reliable prediction about whether it will outperform it on a different model. A Self-Refine setup that works well on one backbone may underperform a single call on another. Validating an orchestration choice on one model is not sufficient before committing to it across a system with multiple backbone options.

What this study does not claim

Orchestration gains are positive in these results. The claim is not that multi-call methods are useless. It is that the cost case for orchestration is much weaker than the literature implies once both sides have been optimized equally. A gain of 4.6 pp at 2 to 4 times the tokens is a real engineering tradeoff, not a clear win. The study covers three orchestration families and three domains. Results may differ in other settings.

What this means
for building with LLMs.

The practical implication is not "abandon orchestration." It is "stop comparing a tuned orchestration stack to an untuned baseline and calling that an evaluation."

1
For teams evaluating orchestration adoption
Invest equivalent optimization effort in your single-call baseline before concluding that a multi-call architecture is worth the added cost. If you tune the orchestration arm and leave the baseline at defaults, you are measuring the effort gap, not the architecture gap. Use a consistent optimization procedure on both sides, then compare accuracy-per-token across each backbone you plan to use in production.
2
For teams already running orchestration
Run the same optimization procedure on a single-call baseline and measure whether the accuracy difference still justifies the token overhead at your production volume. If the gap closes significantly, the value may be in how the orchestrated arm was configured, not in the architecture itself. That is a much cheaper fix.
3
For teams using task difficulty to justify orchestration
This study found no evidence that harder tasks widen the orchestration advantage. If your internal justification for the orchestration overhead is "it will really earn its keep on the hard cases," that assumption is now explicitly unverified in controlled conditions. Test it in your domain before relying on it as a cost justification.
4
For teams switching or mixing backbone models
An orchestration configuration validated on one backbone gives no reliable signal about whether it will win on a different one. Before deploying a multi-call setup across a system with multiple backbone options, or migrating an orchestration strategy to a newer model, re-validate the comparison on each backbone independently.
5
For AI teams producing internal benchmarks
Report accuracy-per-token per backbone as the primary evaluation metric when comparing single-call and multi-call approaches. A single cross-model average hides the backbone interaction this study found. An architecture that looks good in aggregate may be dragged up by one strong backbone while underperforming on another.

Where to go
from here.

Concrete steps for applying this work.

1
Read the paper
Leins, N., Pelleriti, N., Gonnermann-Müller, J., & Pokutta, S. (2026). When Does LLM Orchestration Pay Off? A Controlled Evaluation of Accuracy, Cost, and Task Difficulty. arXiv:2608.00685.
2
Audit your current evaluation setup
Check whether your single-call baseline for any orchestration comparison has received the same prompt-engineering and configuration effort as the orchestrated arm. If it has not, that is the first thing to fix before drawing any conclusions from the comparison.
3
Measure cost, not just accuracy
Compute accuracy-per-token for each method and backbone you are comparing. A 4.6 percentage point accuracy gain at 3x the token cost is a different engineering decision than the same gain at 1.5x. The cost-accuracy tradeoff should be explicit before any architectural commitment.
4
Stratify your internal test set by difficulty
If you have not stratified your internal evaluation items by difficulty level, doing so lets you test whether orchestration gains concentrate on harder tasks. If they do not, that changes the cost justification significantly.
5
Re-validate on each backbone you plan to use
Given the strong method-by-backbone interaction, treat any orchestration validation as backbone-specific. Budget a re-validation run for each backbone before it enters production, rather than assuming an evaluation result transfers.