The control tower for AI spend

Know what your AI actually costs. Before the invoice does.

SpendTower shows every token and dollar across Claude, OpenAI, and Gemini, forecasts where the bill is heading, and keeps a human in the loop before spend runs away.

Founding partners get a full year free, a direct line to the team, and a say in the roadmap. A small cohort. Early access opens soon.

Read only
We connect through provider usage and billing APIs. We never sit in your request path, and we do not see your prompts.
Live in an afternoon
Connect a provider and reach a populated dashboard the same day, in under an hour of setup.
Reconciled
Our totals land within 2% of each provider's own invoice, so finance can trust the number.

The problem

Teams move fast. The bill moves faster.

Leadership wants everyone using AI. So keys get handed out, spend scatters across a dozen provider consoles, and nobody owns the total. Then the invoice arrives.

01

Spend is opaque and lumpy

There is no single trustworthy number for what you are spending this month, let alone per team or per product.

Finance is guessing
02

There is no forward view

An eight thousand dollar month quietly becomes a forty thousand dollar month, and the first anyone hears of it is the bill.

No time to react
03

Model choice goes unmanaged

Engineers reach for the biggest model by habit, on workloads where a smaller one would score within a point or two.

Paying for headroom
04

There are no guardrails

A bad deploy or a runaway retry loop can burn a month of budget overnight, and nothing flags it until reconciliation.

No way to intervene

How it works

Live in an afternoon, with no risk to production.

Three steps, none of which touch the code path your customers depend on.

Connect read only

Link each provider with a read-only key scoped to usage and billing. Nothing is installed, nothing proxies your traffic, and no prompt content ever reaches us.

About 10 minutes per provider

See the number, and the forecast

Unified dollars and tokens across every provider, sliced by team, model, and key, with a projection of where the month actually lands.

Populated the same day

Set thresholds and route them

Budgets per team, alerts to email and Slack, and an escalation ladder that ends with a person approving, never an automatic switch.

Ongoing, once configured

What it does

Four jobs, one control tower.

Each one exists to answer a question the budget owner cannot answer today.

01 / Attribution

See it all, attributed to a team

Every token and dollar across every provider, mapped to the team, project, and key that spent it. This is the number most organisations cannot produce today, and it is usually the first thing that changes a conversation.

Our totals reconcile to within 2% of the provider's own invoice.

Spend by providerAug, month to date
Claude 47% OpenAI 31% Gemini 19%
Top teamsSpend MTD
Search team Claude Opus, 9.8M tokens $14,900
Support bot GPT-4o mini, 4.1M tokens $4,120
Data pipeline Gemini Flash, 3.6M tokens $3,880
02 / Forecasting

See it coming, with days to spare

Run-rate projection to month end, budget versus forecast per team, and days to limit. A three times spike gets flagged within a day, not discovered during reconciliation three weeks later.

Every budget overage is warned with at least five days of runway.

Actual versus forecast42% over budget
Actual to date Forecast $61,200 Budget $43,000
03 / Savings

Spend less without flying blind on quality

Anyone can tell you to use a cheaper model. SpendTower shows what the swap costs you in accuracy first, backed by quality signals, so the call is yours and it is an informed one.

Identified versus realised savings, tracked against what SpendTower costs.

RecommendationsCost / accuracy
Search: Opus to Haiku high-volume classification -41%
accuracy -2.1%
Docs: GPT-4o to 4o mini retrieval, short answers -63%
accuracy -3.4%
Support: cache repeated prompts 38% are near duplicates -22%
accuracy 0%
Savings identified this month$9,240
04 / Control

Stay in control, without a kill switch

Threshold rules escalate by severity and route to the people who can act. At the top of the ladder a person reviews and approves. SpendTower never throttles anything on its own.

Who was notified, who approved, and what happened, all in one audit log.

Waiting for reviewSearch team
Search team hit 100% of budget. Recommended: throttle the runaway job to a $500 per day cap. Everything else keeps running.
Awaiting approval from CFO and CTO
Nothing happens until a human says so

Governance done right

Rules trigger. A human decides.

A kill switch that wrongly throttles production is worse than the overspend it prevents. So SpendTower raises the flag, routes it to a person with the context to judge it, and waits. Automatic enforcement is something you can opt into much later, never something we turn on for you.

There is no automated kill switch, by design.
80%Notify the owning team
90%Notify the CFO, warn the team
95%Notify CFO and CTOApproval required
100%Throttle runaway jobHuman approves

Why SpendTower

The pieces exist. Nobody assembled them for the budget owner.

Each neighbouring category does part of this well. None of them was built for the person who signs off on the spend.

Capability comparison between LLM gateways, observability tools, cloud FinOps platforms, and SpendTower
Capability CategoryGateways CategoryObservability CategoryCloud FinOps CategorySpendTower
One spend number across providers
Attribution to named teams
Month-end forecast versus budget
Right-sizing with the accuracy tradeoff shown
Escalation with human approval
Built for finance, not only engineering
Core to the category Partial or varies by vendor Not what it is for

This is our read of the category as of August 2026, and individual products vary a great deal. If we have a neighbour wrong, tell us and we will fix it.

Questions

The things buyers ask first.

Mostly about access, risk, and what we can actually see. Short answers here, the longer list on the FAQ page.

Do you see our prompts or completions?

No. We read usage and billing metadata through each provider's API: model, token counts, cost, key, and timestamp. Prompt and completion content is never requested and never stored.

Does SpendTower sit in our request path?

No. We are read only by design, so an outage on our side cannot affect your production traffic. An optional inline gateway for real-time enforcement is on the long-term roadmap as something you opt into, not part of the core product.

Will it ever throttle production on its own?

No. Rules escalate and notify, and a person approves before any action is taken. There is no automated kill switch, and that is a deliberate product decision rather than a missing feature.

Which providers do you support?

Anthropic, OpenAI, and Google are first, through their usage and billing APIs. Longer-tail providers follow once the first three are solid. If you are on something else and spending seriously, tell us and it will weigh on the order.

What does it cost?

Pricing is not published yet, and we would rather not invent a number before founding partners have shaped the product. Founding partners get a full year free. Our intent is that realised savings run at several times whatever SpendTower ends up costing.

Founding partners

Get a full year free. Help shape what gets built.

We are taking on a small cohort of founding customers. In return we want your honest feedback, including the parts we will not enjoy hearing.

Apply as a founding partner
  • A full year of SpendTower at no cost
  • A direct line to the people building it
  • Real influence over what ships and in what order
  • Your connectors and edge cases prioritised

Early access

Join the waitlist.

Be first in when early access opens, and get a say in the roadmap. No newsletter, no drip sequence, just the launch.

Prefer to talk first? Become a founding partner. Read how we handle your details in the privacy policy.