Token Cost ROI: 50 agency workflows measured at scale.

Tokens were always the wrong unit. The agency conversation in 2024 and most of 2025 obsessed over $/1M tokens — a metric that ignores whether a run produced a shippable artifact, and that mis-prices cheap workflows as wins and expensive ones as losses. The right unit is cost-per-successful-task against attributed revenue lift, and that is what this post measures across 50 production-instrumented workflows.

The dataset covers a six-month window through April 2026, spanning 15 marketing workflows, 20 dev workflows, and 15 hybrid workflows (lead enrichment, competitor analysis, client reporting). Every run was instrumented through OpenTelemetry traces, with per-run token spend reconciled monthly against provider invoices and revenue lift attributed through a 30-day forward window against a matched manual baseline run by a senior strategist.

The headline numbers will surprise even practitioners who track their spend carefully. Median per-run cost ranges from $0.07 (email triage) to $12.40 (long-context competitor analysis), with one outlier run at $87. Median ROI ranges from 11.4× (SEO audit) to 1.6× (client report). And prompt caching cuts cost-per-successful-task by 38–72% on workflows with repeated input — meaning the ROI table reorders once you turn caching on.

Through 2024 and most of 2025 the agency-AI conversation was anchored to $/1M tokens. Vendors competed on it, blog posts ranked models by it, and procurement teams built spreadsheets around it. The metric is fine for capacity planning and useless for unit economics — because it measures the cost of a generated string without asking whether the string was useful.

The right unit is cost-per-successful-task: total spend divided by the number of shippable outputs the workflow produces. That number can be ten times higher than the per-token cost would suggest (when failure rates are high) or ten times lower (when caching is aggressive and the workflow is repeated). Both distortions matter, and neither shows up if you only track $/token.

The corollary is that ROI — attributed revenue lift divided by cost-per-successful-task — is the only honest comparison across workflows. A $0.07 email-triage run and a $4.20 SEO-audit run are not comparable on cost; they are comparable on ROI. The audit, at 11.4×, is dramatically more valuable than the triage at 2.1×, even though it costs sixty times more per run.

The dataset is 50 production workflows running across agency engagements over a six-month window through April 2026. Workflows split 15 marketing, 20 dev, and 15 hybrid. Every run produced an OpenTelemetry trace including model, input tokens, output tokens, cached-read tokens, reasoning effort tier, latency, and a structured success label assigned by a senior reviewer within 48 hours of the run.

Cost was reconciled monthly against provider invoices — Anthropic, OpenAI, Google, and a small DeepSeek allocation — to catch the gap between rack-rate token math and actual billed amounts (cache credits, volume discounts, and committed-use rebates all change the number). Revenue lift was attributed through a 30-day forward window against a matched manual baseline run by a senior strategist on the same client account in the prior quarter, with 95% confidence intervals on every median.

The key methodology choice is the success label. A run that generated tokens but produced output the strategist would not ship counts as a failed run with full cost charged and zero revenue attributed. This is what makes cost-per-successful-task different from per-run cost: failures are amortized across the successes that pay for them.

Median per-run cost spans almost three orders of magnitude — from seven cents for an email triage to over twelve dollars for a long-context competitor analysis, with one observed outlier at $87. The bar chart below shows the median for each workflow class after caching, model mix, and reasoning-effort tuning have been applied.

Two patterns jump out. First, the high-volume workflows (enrichment, triage, ad-copy iteration) sit at the bottom because they run thousands of times per month and any cost above $0.50 per run becomes economically untenable. Second, the high-cost outliers (competitor analysis, SEO audit) are exactly the workflows where humans previously spent five to ten hours per output — so even at $12 per run the cost-vs-labor swap is dramatic.

The $87 single-run outlier is instructive. It was a competitor analysis on a B2B portfolio with twelve competitor sites, run on Claude Opus 4.7 at full 1M context with extended thinking enabled. The per-run cost looks alarming until you compare it to the seven hours of senior strategist time that the manual baseline took — at $250/hour blended rate, a $87 spend that returns the same artifact in 40 minutes is roughly a 20× cost reduction.

ROI here is attributed revenue lift over a 30-day window divided by cost-per-successful-task. The denominator includes the cost of failed runs amortized over the successes; the numerator counts only revenue plausibly attributable to the workflow output (new retainer expansion, won ad spend, billable deliverable hours replaced). The table flips the order of the cost ranking — the cheapest workflows are not the most valuable.

The pattern is clean: workflows that unlock new revenue (audits, enrichment, outreach) post the highest ROI. Workflows that replace operational cost without unlocking revenue(triage, reports) sit at the bottom. This is intuitive in retrospect but it inverts the usual cost-first prioritization — agencies that build their AI roadmap by "cheapest first" end up shipping the lowest-ROI workflows first.

Prompt caching is the single highest-ROI infrastructure investment for an agency AI stack. The savings on cost-per-successful-task — not raw $/token, the right unit — range from 0% on one-shot workflows where the input never repeats, to 72% on workflows that re-process a stable corpus across many runs. The four numbers below cover the workflows where caching matters most.

The shape of these numbers maps directly to a workflow design rule: identify the stable layer of your prompt (corpus, brand voice, prior context) and the variable layer (the per-run question), then push the stable layer into the cache. Workflows that fit this pattern see double-digit percentage cost reductions; workflows that do not (one-shot enrichment, single-pass triage) see zero benefit and should be optimized through model selection and reasoning-effort tier instead.

Note the second-order effect on the ROI table. Client reports at 1.6× ROI are the worst performer in the raw table — but with caching applied, cost-per-successful-task drops 47% and ROI rises to roughly 3.0×. The ROI ranking is sensitive to caching discipline; agencies that have not yet implemented caching are looking at a distorted picture of which workflows are working.

No single 2026 frontier model wins every workflow. The dataset shows clear winners by workflow class — driven by latency, context-window economics, multimodal capability, or open-weight cost floors. Agencies running a single-model stack are leaving 30–50% of margin on the table compared to a routed mix.

The routing logic is simple in principle, harder in practice: workflow class maps to model class, and the stack should route at the workflow boundary rather than at the user-prompt level. Agencies that build a routing layer once — even a simple workflow-to-model dictionary — capture most of the available margin. Agencies that route per-prompt or per-user end up with inconsistent quality and unnecessary infra complexity.

The dataset surfaces four recurring patterns that inflate cost-per-successful-task without lifting ROI. Each one is fixable inside a single sprint, and the combined savings on a typical mid-size agency stack run 40–60% of monthly AI spend.