Every AI cost tool estimates from a price sheet. Provider invoices don’t work that way — cache multipliers, reasoning tokens, mid-month price changes, usage revised after the fact. ToTheCent ties every request to the line it became on your provider invoice. Then it proves what your AI actually costs, wastes, and earns.
Anthropic spend is reconciled to its prepaid credit invoices and account balance; other providers are priced from published rates — where reconciliation succeeds, the residual stays within half a cent per line of the provider’s cost report; where it can’t, ToTheCent tells you exactly why instead of hiding it (methodology).
Read-only. Token counts and hashes only — never your prompts.
ToTheCent’s core is deterministic SQL — zero LLM anywhere in the numbers path. Every figure traces to an invoice line or is explicitly labeled MODELED with its named assumption. Never one blended ROI number.
Per-request true cost matched to provider invoices — and every gap explained, never averaged away.
The data model is FOCUS-native, so AWS, Azure, and GCP bills are designed to join the same reconciliation engine. That is not live today: for cloud, ToTheCent at most reads the closed-bill total from an uploaded FOCUS export — no cloud allocation, no cloud waste findings. What is live is AI-provider spend: Anthropic reconciled to prepaid credit invoices and balance, other providers priced from published rates.
The gross-margin-per-feature answer investors ask for.
Invoice → general ledger → margin bridge; measured savings separated from modeled forecasts, exceptions shown, never dropped. NPV, IRR, payback — with optimism explicitly haircut, never baked in.
Reconciled to the cent at the finest grain your provider’s invoice discloses, with any residual shown — never hidden.
Cost attribution and showback per tenant, feature, and customer. Your system of record for AI spend.
One dollar-ranked list across your AI bill: model right-sizing, caching, context bloat, retries, batch. Anti-double-counted, each shipped as a ready-to-apply config diff plus an eval plan — hand it to your engineer or your coding agent, run the evals, merge. ToTheCent never touches your production path; you approve every change. And after you ship a fix, ToTheCent reprices your actual workload against the next invoice — so every claimed saving is measured, not asserted.
Most AI initiatives can’t demonstrate ROI — not because there is none, but because nobody defined benefits, baselines, and owners before shipping. ToTheCent’s guided interview does exactly that: benefit classes and conversion paths, costs beyond inference, hurdle rates, ramp phasing, a skeptic pass that challenges optimistic assumptions against reference-class priors. Then actuals flow in monthly from reconciled cost, and MEASURED vs MODELED variance is tracked — so six months in, you know whether the case held, not just whether it was pretty.
Two artifacts, one engine, exportable in a day:
Verified Cost Report. What you actually spend on AI — every dollar tiered by how it’s verified: reconciled to the invoice, invoice-verified, management-declared, or uncovered. Coverage sums to 100% — we tell you what we couldn’t verify, too. Includes deterministic stress-replay: your real workload repriced under +25% provider prices, caching off, 5–10× load — same SQL, same hash, run twice.
Value Case. What it returns — NPV, IRR, payback with explicit bias adjustment, benefit targets vs measured actuals month by month, and a closed loop on every optimization: each shipped fix counterfactually repriced against the next invoice, so “we saved $X” is a reconciled number, not a slide.
A boutique cost review runs €5,000–15,000 and takes 2–4 weeks — opinions in a PDF you can’t re-run. AI-specific or enterprise FinOps work runs well past €30,000, often 4–8 weeks, with the same limitation. ToTheCent’s report is generated from your reconciled data, reproducible bit-for-bit, in a day, on a plan from $250/month.
Cache writes billed at 1.25×–2×. Reasoning tokens billed as output. Prices changed mid-month. Batch and cache discounts stacking. Usage revised after the month closes. Price-sheet math silently drifts from the bill — and ToTheCent’s effective-dated rate engine catches the drift before you see it on the invoice.
Bring the telemetry you already have — Langfuse, OpenTelemetry GenAI, or our thin SDK wrapper. ToTheCent ingests token counters and metadata (never message content), reconciles against your usage API and invoice, and puts every number behind a SQL query you can inspect. Narrative reports are generated by an EU-resident model by default, with every number cited from SQL and hard-blocked if grounding fails.
The FinOps playbook — attribution, tag coverage, anomaly detection, rate-drift monitoring, savings ledgers, business-case governance — built into the product instead of a hire you can’t make yet. FOCUS-aligned data model, FinOps-standard KPIs, and the discipline enforced by the engine, not by a slide deck.
One engine, one methodology — live for your AI-provider invoices, designed for your cloud bill next.
Dashboards estimate. ToTheCent reconciles. If a number doesn’t tie to an invoice line, it’s labeled modeled — or it doesn’t ship.
You’d get estimates. Reconciliation requires an effective-dated rate engine, a residual-allocation waterfall, and restatement handling — and a build-it-yourself spreadsheet won’t survive diligence.
Price, speed, and reproducibility. A boutique review runs €5–15k and takes 2–4 weeks; AI-specific or enterprise work runs well past €30k and 4–8 weeks — and both deliver opinions you can’t re-run. ToTheCent’s report is generated from your reconciled data by a deterministic engine — every number traceable to SQL and invoice lines, reproducible bit-for-bit, on a plan from $250/month (pricing), in a day. When a number is challenged, you re-run it instead of scheduling a call.
Not yet. Cloud FinOps is designed for, not live: the data model is FOCUS-native, but today ToTheCent at most reads the closed-bill total from an uploaded FOCUS export. It does not yet parse CUR or BigQuery exports, allocate cloud costs to teams, or detect cloud waste. What ToTheCent reconciles today is AI-provider spend.
No. Token counts, hashes, and business metadata only.
Never. Read-only, off-path. Remediations ship as config diffs with eval plans; you apply them.
No one honestly can. We show dollar-ranked opportunities grounded in reconciled cost — and after you ship a fix, we counterfactually reprice your workload against the next invoice, so the savings ledger tracks what actually landed in your P&L.
On EU infrastructure. Report narratives are written by Anthropic Claude by default, from aggregated figures only — never your prompt content. An EU-hosted Mistral option is available on request.
Send one invoice. Watch it reconcile.