Summary: Token prices are public and falling. Agent budgets still explode, because the cost that matters is not the price per million tokens; it is how many tokens your architecture burns per business decision. This post walks through measured numbers from a July 2026 study: what the same workload costs as a single structured pass, as an iterative agent, and as a voting ensemble, and what that means for anyone budgeting an agent program in 2026.
The number that matters: cost per decision
Every vendor quotes cost per million tokens. No CFO cares. The budget line that matters is cost per decision: what it takes to review one loan file, adjudicate one claim, or evaluate one compliance case, end to end.
Cost per decision is where architectures separate. The same model, at the same public price, can cost 3x to 10x more per decision depending on how the system uses it. That variance, not model pricing, is why 80% of companies exceed their AI cost forecasts.
What the measured study found
A July 2026 measured study ran the same multi-document evaluation workload in three configurations on live models, counting every token from API usage records rather than estimating:
- Single structured pass: $1.34 per evaluation cycle. Context loads once. The model evaluates against a fixed rubric with constrained output.
- Growing-context agent: $4.64 per cycle. The agent re-reads its own history at every step, replaying 3.6x the input tokens of the single pass for the same verdicts.
- Voting configurations: $10.63 to $15.27 per cycle. Running 2 or 3 independent passes plus a verification step to chase reliability multiplies the whole cost structure.
Three findings travel beyond this workload. First, input replay is the dominant waste: iterative agents pay repeatedly for context they have already read. Second, caching helps asymmetrically: cached input reads bill near a tenth of the base rate, but output tokens are never cached, so chatty architectures stay expensive even with perfect cache discipline. Third, buying reliability with redundancy has a floor: voting converges to a borderline error rate near 10%, so you can triple spend without reaching production-grade accuracy. We covered that trap in depth in the voting economics analysis.
The architecture decision is the budget decision
Map the three configurations onto the platforms you are evaluating and the budget writes itself.
Reasoning-loop agents (the ReAct pattern and its descendants) are the growing-context configuration: every step replays context, every failure retries, and per-decision cost varies with document complexity, retry luck, and reasoning depth. You can measure the average; you cannot forecast the specific case.
Compiled execution plans are the single-pass configuration generalized: the plan is fixed before runtime, deterministic checks run as code at zero model cost, and model calls are constrained to the steps that need judgment. Per-decision cost becomes a number you can put in a budget, because the same case type consumes the same tokens every time. This is how MightyBot executes workflows, and it is why the platform holds a 10x token-efficiency advantage at production volume.
Budgeting rules for 2026
- Price the decision, not the token. Ask every vendor: what does one completed case cost at my volume, and what is the variance?
- Ask what happens on failure. Retries are the hidden multiplier. A 20% retry rate on a context-heavy step can double a workflow’s real cost.
- Model the years, not the month. Model prices improve roughly 25% annually, but that credit accrues to every architecture equally. The architectural multiplier persists.
- Run the comparison yourself. The AI Agent ROI calculator models build-versus-buy cost with the measured data above, including cache behavior, quality strategies, and 3-year trajectories.
Token prices will keep falling. Architectures that waste tokens will keep outspending architectures that do not, by the same multiple, at any price.
