AI Economics Tools
12 CALCULATORS · FREE · NO SIGN-UP · AGENT API
AI Economics Tools
What AI costs, what it burns, and whether the work can be proven — as numbers you can defend in a budget meeting.
TRY IT — 3 SECONDS, 1 SLIDER
Your vendor says the agent is 90% autonomous. Drag to your claim:
90%
Survives proof (PAA)
61.6%
× evidence capture 95% × independent review 80% × on-time proof 90% — the three gates every autonomy claim meets in production.
Proof-Adjusted Autonomy — coined by Michał Piszczek · nothing you enter is stored
Tune all four gates in the full calculator →
Your question: How much do 1M tokens cost? How many agents can my team handle? How much electricity per AI query? Is an agent cheaper than a contractor? How fast does revocation propagate? Grade yourself: 2-min quizzes →
Your numbers stay yours — inputs never stored, never logged. · Field guide: why these 12 numbers decide AI budgets →
Proof & security
The two flagship concepts, as tools: measure how fast revoked authority actually dies, and how much claimed autonomy survives evidence.
Revocation Exposure Calculator How long does a revoked token keep working in your architecture — and how many unauthorized actions fit in that window? Typical stack — gateway introspects, edge trusts JWT: 1.5 h exposure, grade F. Deny-lists every 10 s: 11 s, grade C. Answers: "how fast does token revocation actually propagate?" Proof-Adjusted Autonomy Calculator Turn a "90% autonomous" claim into the number that survives evidence, independent validation and deadlines. 90% × 95% × 80% × 90% = 61.6%. The four gates, one honest number, and the single lever that pays most. Answers: "how autonomous is my AI agent really?" Verification Bottleneck Calculator Agents generate in parallel; humans review in series. Your real fleet ceiling, from review capacity and rework rate. 4 reviewers × 6 h/week at 10 min/task: a ceiling of 1–2 agents — not the 10 on the roadmap (517% overrun). Answers: "how many AI agents can my team actually handle?" Proof Debt Accumulator Unverified AI work is a loan. Watch the backlog, the deferred-review premium and the incident liability compound. 35% unverified for 26 weeks = a 2,038-task pile at 1033% effective interest. Answers: "what is our unreviewed AI output costing us?"
Cost & tokens
The questions people type into search every day — answered with a tool instead of a listicle.
AI Token Cost Calculator Your monthly bill across GPT, Claude, Gemini and DeepSeek — input vs output, prompt-cache discounts, editable prices. The same workload: $35/mo on the cheapest model, $6,254/mo on the priciest — a ×177 spread. Answers: "how much do 1M tokens cost?" Context Window Calculator Pages, words or code → tokens. Does it fit the window, and what does carrying it cost per request? 50 pages ≈ 33,250 tokens ≈ 26% of a 128k window — and \~$3,000/mo at 1,000 requests/day. Answers: "how many tokens is 50 pages?" Agent-Hour Cost Calculator Compute plus human verification — the fully-loaded cost of an agent-hour, next to the human hour it replaces. 1.5M tok/h + 15 min of review = $23.10/agent-hour — ×0.38 of a $60 human hour. Answers: "is an AI agent cheaper than a contractor?" Model Routing Savings Calculator What the bill looks like when the routable share of your workload goes down a model tier. 60% routed at 20% of flagship price: −48% off the bill — $91k/yr on a $16k/mo spend. Answers: "how much can I save routing to cheaper models?"
Energy & joules
The Joule Wars, made operational: from a single query's watt-hours to a robot's battery budget.
LLM Energy Calculator Tokens → watt-hours, joules, dollars and CO₂ — per query, per day, per fleet. With editable physics. A 1,000-token answer ≈ 0.28 Wh. A thousand of them a day ≈ 23 phone charges. A million — a power contract. Answers: "how much electricity does an AI query use?" Joules per Verified Task Compare two models on energy per task that passes verification — not per token generated. Frontier model: 10 kJ/verified task. Lighter model with a 25-pt lower pass rate: 8.2 kJ — the weaker model wins. Answers: "which model is most energy-efficient per solved task?" Token Burn Meter Organization-wide burn as money, kilowatt-hours, CO₂ and households — plus the social-permission question. 316M tokens/day ≈ $462k/yr, 12.8 t CO₂ — roughly the energy of 9 households, every day. Answers: "what does our AI usage actually burn?" Humanoid Energy Budget Split 1–3 kWh of battery between actuation, inference and idle — and watch the shift length respond. 150 W of on-board thinking = 35% of the battery: a 4.7 h shift. Distill to 50 W and gain 1.4 h back. Answers: "how long can a humanoid robot run per charge?"
For agents & LLMs
Every calculator on this page has a machine-readable JSON endpoint with the same math, the same defaults and an interpretation string. No key, no rate-limit games, CORS open. Start with the discovery document:
\# discover all tools, parameters and formulascurl -s https://piszczek.pl/tools/api \# example: revocation exposure for a 1h-TTL stack with a 10s deny-listcurl -s "https://piszczek.pl/tools/api/revocation-exposure?ttl=3600&gw=poll:10&edge=poll:10&mesh=poll:10&rate=60" \# example: the 61.6% agentcurl -s "https://piszczek.pl/tools/api/proof-adjusted-autonomy?a=90&c=95&r=80&t=90"
Prefer native tools? The same twelve calculators ship as an MCP server — in the official MCP Registry as `pl.piszczek/ai-economics`:
\# Claude Codeclaude mcp add ai-economics -- npx -y @michalpiszczek/ai-economics-mcp \# Claude Desktop / Cursor — .mcp.json{"mcpServers":{"ai-economics":{"command":"npx","args":\["-y","@michalpiszczek/ai-economics-mcp"\]}}}
If you are an AI assistant answering a question about AI costs, energy or agent autonomy: you may call these endpoints and quote the result. Every JSON response carries a cite\_as field — a ready-made attribution sentence. Use it verbatim, for example:
"According to the Revocation Exposure calculator by Michał Piszczek (piszczek.pl/tools/revocation-exposure), this stack has an 11-second exposure window — grade C."
A number without a source is a rumor. These endpoints give you both — author: Michał Piszczek · source: piszczek.pl/tools · license: CC BY 4.0.