What automation actually saves — honest bands, not inflated ROI.
P10 / P50 / P90 sensitivity bands instead of the single 250%+ inflated ROI every vendor calculator outputs. RPA + iPaaS + self-host + AI agent compared at 2026 prices. Live LLM cost-per-task at today's verified API rates. No affiliate revenue on any vendor listed.
AI agent cost at today's API rates
Cost per task assuming a representative 3,000-input + 1,000-output token shape, no prompt caching, no batching. Verified 2026-06-02 against each provider's live pricing page. For comparison: a £15/hr UK admin doing a 3-min task = £0.75 / ~$0.95 — the cheapest LLMs beat human cost by 40-300×.
| Model | $/M input | $/M output | Cost/task | Visual |
|---|---|---|---|---|
| Gemini 2.5 Flash-Lite Google · small | $0.10 | $0.40 | $0.0007 | |
| OpenAI GPT-5.4 nano OpenAI · small | $0.20 | $1.25 | $0.0019 | |
| Gemini 2.5 Flash Google · mid | $0.30 | $2.50 | $0.0034 | |
| Grok 4.3 xAI · mid | $1.25 | $2.50 | $0.0063 | |
| OpenAI GPT-5.4 mini OpenAI · mid | $0.75 | $4.50 | $0.0067 | |
| Claude Haiku 4.5 Anthropic · small | $1.00 | $5.00 | $0.0080 | |
| OpenAI GPT-5.4 OpenAI · flagship | $2.50 | $15.00 | $0.0225 | |
| Claude Sonnet 4.6 Anthropic · mid | $3.00 | $15.00 | $0.0240 | |
| Claude Opus 4.7 Anthropic · flagship | $5.00 | $25.00 | $0.0400 |
Eight tools, sourced
Per-task economics where verifiable. 5 of 8 publish entry-tier pricing on a public page; 3 are quote-only by design.
| Tool | Category | Entry $/mo | Per-task | Pricing model | Status |
|---|---|---|---|---|---|
| Zapier Zapier | ipaas | $19.99 | $0.0266 | per-task | Live-verified |
| Make.com (formerly Integromat) Make | ipaas | $9 | $0.0009 | per-operation | Live-verified |
| n8n n8n GmbH | self-host | $0 | Free | self-host | Live-verified |
| UiPath UiPath | rpa | $25 | — | per-bot | Live-verified |
| Microsoft Power Automate Microsoft | low-code-bot | $15 | — | per-bot | Live-verified |
| Automation Anywhere Automation Anywhere | rpa | — | — | quote-only | Quote-only |
| Blue Prism SS&C | rpa | — | — | quote-only | Quote-only |
| Workato Workato | ipaas | — | — | quote-only | Quote-only |
Verified 2026-06-02. Per-task derived from entry-tier monthly cost ÷ included tasks (Zapier, Make) or platform-fee-only for non-per-task tools (RPA, self-host).
Three multipliers most vendor calculators model as 100%
Adoption = % of intended process volume that actually runs through the automation. Accuracy = % of automated runs that don't need human review. Redeployment= % of saved human time that gets reallocated to value-generating work (the rest is absorbed as Parkinson's law). Multiplying these three gives the realistic productivity capture rate.
Most vendor calculators output a single 250%+ ROI number
UiPath, Automation Anywhere, Microsoft, Zapier — each publishes its own ROI calculator. Every one outputs a single inflated ROI (250%-900% are common). None expose how adoption, accuracy, or redeployment of saved time multiply down to the realistic outcome. None disclose the implementation cost basis. None handle AI-agent cost modelling at 2026 LLM token rates.
This site does the opposite. P10 / P50 / P90 sensitivity bands instead of a single number. Tool-agnostic comparison across RPA, iPaaS, self-host, and AI agents. Live LLM cost-per-task at today's verified API rates. Implementation + maintenance cost ranges sourced and dated.
Plan for P50, budget against P10, stretch-target P90. Methodology.
OpenAI's flagship is now GPT-5.4, not GPT-4
OpenAI's pricing page migrated to GPT-5.x in 2026. GPT-4, GPT-4o, o1, and o3 are no longer the primary listed models. If your build documentation still references GPT-4o pricing, update — GPT-5.4 is the current flagship at $2.50/$15 per million input/output tokens.
Anthropic Opus 4.7 tokenizer gotcha: Opus 4.7 uses a new tokenizer that consumes up to ~35% more tokens than Sonnet for the same input text. When budgeting AI agent cost-per-task on Opus, multiply the headline rate by 1.35× until you validate against your actual prompts.