Measured evidence · July 2026

What multi-agent evaluation actually costs: a measured study

We reconstructed a production-shaped, multi-document quality-gate workload on a fully synthetic corpus, then ran every competing agent configuration on the same model at temperature 0 and counted every token from the API usage records. The result is a like-for-like cost ladder for single-pass evaluation, growing-context agentic evaluation, and self-consistency voting, with and without prompt caching. These measurements power the defaults in our AI Agent ROI & TCO Calculator.

The short answer

On an identical 11-document evaluation workload, a single-pass multi-viewpoint evaluation cost $1.34 per cycle and hit output caps. A growing-context agentic configuration cost $4.64 per cycle, replaying 3.6x the input tokens of a single pass. Pushing agents toward deterministic-grade reliability with majority voting and a verification pass cost $10.63 to $15.27 per cycle, took tens of minutes, and still left roughly a 10% error rate on borderline decisions. Prompt caching compresses these numbers to $0.64, $1.85, and $4.35 to $6.20 respectively, but it cannot change the ordering, because voting multiplies output tokens and output tokens are never cached.

Method

  • Workload. A synthetic corpus of 11 software design documents (about 130,000 characters) with a 58-item quality checklist, requirement traceability, and identifier contracts, generated for this study. No customer data was used.
  • Ground truth. A deterministic rule implementation of the checklist was verified first: a healthy corpus variant returns pass with zero findings, and a degraded variant (one truncated document, one stubbed core document, two broken must-priority traces) returns exactly 5 findings at coverage 0.947. All agent configurations were graded against this ground truth.
  • Measurement. Every configuration ran on the same commercial model at temperature 0. Token counts come from the API usage object of each call, not from estimates. Costs use list pricing: $3.00 per million input tokens, $15.00 per million output tokens; the cached scenario bills cache reads near 0.1x with a small write premium.
  • Configurations. (1) Single-pass evaluation: 12 evaluation viewpoints, each reading the full corpus once. (2) Agentic evaluation: each viewpoint as a 4-turn growing-context agent. (3) Parity attempt: 2 to 3 independent agentic passes with majority voting plus a single-pass verification sweep.

Measured results

Configuration Tokens per cycle Cost per cycle (uncached) Cost per cycle (cached) Reliability observed
Single-pass evaluation (12 viewpoints) 297.9K in / 30.0K out $1.34 $0.64 Outputs hit token caps; misses leave no recorded reason
Agentic evaluation (4-turn agents x12) 1.074M in / 94.8K out $4.64 $1.85 3.6x input replay measured; single-run flip rate ~20% on borderline decisions
Parity attempt (2-3 voting passes + verification) ~2.4M to 3.5M total $10.63 - $15.27 $4.35 - $6.20 Error floors near 10.4% at 3 votes; tens of minutes per cycle; no audit trail

Per-viewpoint single-pass measurements: 24,824 to 24,831 input tokens, 2,500 output tokens (capped). Agentic per-viewpoint measurement across 4 turns: inputs of 10,189 / 20,463 / 28,509 / 30,315 tokens with 7,900 output tokens total. The x12 extrapolation applies a measured per-viewpoint cost to the full viewpoint set.

Why voting cannot buy determinism

Residual borderline-decision error vs voting passes (measured flip rate p = 0.20) 0% 10% 20% Deterministic layer: 0% (same input, same verdict, every run) 1 pass · 20% · $4.64 3 passes · 10.4% · $15.27 5 passes · 5.8% · $24.56 Cost of the agent-based configuration (measured cycle x passes + verification)
Majority voting over independent agentic passes reduces the measured 20% borderline flip rate to 10.4% at 3 votes and 5.8% at 5 votes, at $15 to $25 per cycle, and converges to a floor rather than to zero. Shared blind spots are not reduced at all: a check that silently fails to fire fails in every pass. Deterministic checks return the same verdict for the same input every time.

There is a second structural effect: prompt caching compresses input costs on every configuration, but voting multiplies output tokens, and output tokens are never served from cache. That is why the parity attempt stays 3 to 7 times more expensive than a single pass at any caching level, while still failing reproducibility, coverage certification, and audit-trail requirements.

Measured versus modeled

  • Measured: all token counts (API usage objects), the 3.6x agentic input-replay multiplier, per-cycle costs, output caps, and end-to-end wall-clock times of 35 to 105 seconds for compiled single-cycle runs versus minutes to tens of minutes for multi-pass agentic runs.
  • Modeled and labeled: the x12 viewpoint extrapolation from measured per-viewpoint costs, the voting pass count (2 to 3), the majority-vote arithmetic applied to the measured flip rate, and the cached scenario (perfect within-cycle caching, the best case for the agent side since caches expire in minutes and cross-run reuse is rarely realistic).
  • Corpus: fully synthetic, generated for this study at a production-realistic scale. Ratios travel to other workloads; absolute dollar figures shift with document sizes and volumes. The method is reproducible on your own data.

Frequently asked questions

How much does agentic evaluation cost per cycle?

Measured on a reproduced multi-document evaluation workload on live models: $1.34 per cycle for a single structured pass, $4.64 for a growing-context agentic run, and $10.63 to $15.27 for 2-to-3-vote configurations with verification. With realistic prompt caching those figures drop to roughly $0.64, $1.85, and $4.35 to $6.20 respectively.

Why do agentic runs cost more than single passes?

Input replay. The growing-context agent re-reads its accumulated history at every step, replaying 3.6x the input tokens of a single pass to produce the same verdicts. The waste is architectural, not model-dependent.

Does prompt caching fix agent evaluation costs?

It narrows the gap but never closes it. Cached input reads bill at roughly a tenth of the base rate, but output tokens are never served from cache, so iterative and voting configurations remain output-dominated and cost multiples of a compiled pass.

Does majority voting make agent evaluation reliable?

Partially. A roughly 20% borderline flip rate on single-pass evaluation drops to roughly 10% with 3-vote majority, at roughly 3.3x the model spend. Voting cancels randomness but converges to a borderline error floor; it cannot reach deterministic reproducibility.

Use the numbers

These measurements set the architecture multipliers, cache modeling, and quality-equalization defaults in the AI Agent ROI & TCO Calculator. Model your own volumes, labor costs, and caching assumptions there, or read the Build vs Buy guide for the decision framework.