The cost and risk math

The simulator's baseline configuration has an API bill of five dollars. Its true monthly cost is $2,315. Nothing in that gap is exotic: it is seven silently wrong numbers at $330 of analyst cleanup each. The most expensive line in an AI budget has never once appeared on an AI invoice.

Tonight's discipline is simple to state: price the whole stack, not the meter. The meter reads tokens. The stack includes the eval set someone maintains, the monitoring someone watches, the checking someone does, and the cleanup someone gets pulled into. A route is not cheap because its meter reads low. It is cheap when the whole stack reads low - and the deterministic tier keeps winning precisely because its stack is almost all zeros.

Watch the simulator's arc across the sessions: $2,315 at baseline, $225 after triage, $224 on the free tier, $62 with classic ML, $176 with GenAI fit - and then $416 when full costing arrives. The last step goes UP, and it is the most trustworthy number in the whole table, because it is the first one carrying evals, monitoring, and upkeep on every probabilistic route.

A budget that only ever goes down is usually a budget that stopped counting something. The routed config's $416 against the baseline's $2,315 is still better than 5x cheaper - with 10 out of 10 answers trusted instead of 0. That is the two-number story worth telling: the honest number is bigger than the bill, and dramatically smaller than the chaos it replaced.

Five model rates and one query rate cover everything Cadence runs. All figures as of Sept 2026, from the official pricing pages:

Two published discounts change real budgets: the Batch API takes 50 percent off anything that can wait, and cache hits price input at 0.1x - a repeated system prompt costs a tenth of a fresh one. And one quiet tax: newer tokenizers produce roughly 30 percent more tokens for the same text across generations, so a budget built on sticker rates alone runs hot.

Transcript action-item extraction, on Sonnet 5. About 660 runs a month, each around 9,000 tokens in and 500 out. That is 2.3 cents of tokens per run - roughly $15 a month of raw tokens, and with the 30 percent tokenizer headroom budgeted in, the line Cadence carries is about $20 a month. For themed, captured commitments out of every meeting the product records.

Personalized onboarding emails. Only 30 runs a month, but each one reads the customer's full context: about 90,000 tokens in and 300,000 out across drafting and revision. That is 2.7M input ($5.40) plus 9M output ($90) - about $95 a month, the biggest GenAI line in the budget, and worth it: it is the one workload where every generated word is the deliverable.

Every probabilistic route carries the full stack. Every deterministic route skips almost all of it:

• Layer 1 · The bill. Tokens and queries, from Part 1's rate card. The only layer with an invoice.
• Layer 2 · Evals and monitoring: $40 per probabilistic workload per month. An eval set someone maintains and a dashboard someone watches. A probabilistic route with no evals budgeted is a route that has not been approved - it has just not failed loudly yet. Cadence's six probabilistic workloads carry 6 x $40 = $240, which is exactly the gap between the $176 bill and the $416 true cost.
• Layer 3 · Verification labor: $220 a month whenever checkable answers flow through an LLM. If a human must re-derive the number to trust it, the route is paying twice for one answer - which is the accounting argument for never sending checkable work there in the first place.
• Layer 4 · Cleanup: $330 per silently wrong number - six analyst-hours of finding it, tracing it, and repairing what it touched. Seven of these made the baseline $2,315 while its bill read $5.

The stack scales up. Goldman Sachs' 2024 report "Gen AI: Too Much Spend, Too Little Benefit?" is the same accounting argument at macro size, and it is a genuine debate inside one document. Jim Covello, the skeptic, argues that roughly $1 trillion of AI capex has "little to show for it so far" and that AI is not designed to solve the complex problems that would justify the costs. Daron Acemoglu, in the same issue, estimates under 5 percent of tasks profitably automatable and around 0.9 percent GDP impact over a decade.

And in the SAME report, Goldman's own analysts - Briggs, Rangan, and Sheridan - argue the bull side. Quote it as a debate or do not quote it at all: a board paper citing only Covello is as unbalanced as one citing only the bulls. The lesson a router takes from it is neither camp's headline: the spend is real, the returns are concentrated, and honest stack-level accounting is how you land on the right side of the concentration. Cadence's $416 against $2,315 is the same story with fewer zeros.

The figure everyone has heard - 2.9 Wh per ChatGPT query, ten times a Google search - traces to 2024-era estimation (EPRI lineage), built before anyone with production telemetry had published. It has since been superseded by measurement. Google's August 2025 study of live production traffic puts the median Gemini text prompt at 0.24 Wh, 0.03 grams of CO2e, and 0.26 milliliters of water - and reports per-prompt energy down 33x in twelve months. Epoch AI's independent estimate lands in the same neighborhood: a typical GPT-4o query around 0.3 Wh.

Two honest qualifiers travel with the good news. These are medians for text prompts - median is not mean, and agentic or long-context workloads run far heavier than the typical chat turn. And the 33x trend means any number, alarming or reassuring, goes stale fast: quote it with its date or not at all.

Here is why an energy footnote earns a section in a costing session: quoting the stale 2.9 Wh figure in a board paper is the same failure as quoting "95 percent of pilots fail" without its learning-gap caveat. Both numbers were once defensible. Both acquired better evidence. Both keep circulating because they are vivid, and a leader who repeats either one unrefreshed has told the room something false while citing a real source.

The hygiene that fixes both is identical: prefer measurement over estimation when measurement exists, carry the date on every number, and carry the distributional caveat - median is not mean, a pilot statistic is not a capability statistic. A routing playbook is only as credible as the sloppiest number quoted in its defense.