Request a note
AI cost testing · UK

Switch to a cheaper model only when the test says you can.

We build the test from your own examples and agree the pass mark in writing first. If the cheaper model passes, you switch. Thirty days later, we test again.

Request a free written note →

6 questions · written reply within 2 working days · no call

Why one good answer proves almost nothing

The maths

The more of your examples a cheaper model passes, the lower the failure rate it could still be hiding. It takes 59 clean passes to get under 5%.

100% 50% 0 1 5 31 59 examples tested, all passed 5% line 95% after 1 45% after 5
1
pass: could still fail up to 95%
5
passes: up to 45%
59
passes: under 5%

Worst-case failure rate that can still be ruled out at 95% confidence, when every example tested has passed.

How it works

The method
01

Tell us how you use AI

Six questions about your spend band and what the AI does. Redacted bill totals are optional.

02

Get a written noteFree

The likely cause, a rough save range if your models and volumes are clear, and the price.

03

We build the test

Examples from your real traffic, your current model as the baseline, a cheaper candidate, a written list of what each answer must get right, and a pass mark agreed in writing before anything runs.

04

Switch only if it passes

Then a 30-day check. If quality slips, you roll back.

Your real examples 59+ · hard cases shown separately Current model what you run today Cheaper model the candidate Every answer checked errors double-checked PASS MARK, AGREED IN WRITING FIRST PASS · switch FAIL · stay

What you get, and what you don't hand over

The terms
Before you pay anything

What the free note looks like

Every note has the same five parts. You decide from a page, not a sales call.

  1. Your setup. The spend band and tools you told us, so you can check we read it right.
  2. Likely cause. One sentence on where the money probably goes, written as something to test, not a finding.
  3. Rough saving. A range only if you named your models and rough volumes. If not, we say so: sizing it is what the diagnostic is for.
  4. What the £250 diagnostic would contain. Listed in full, so you know exactly what you would be buying.
  5. Price and next step. What we would need from you: a redacted 30-day usage export. Never keys or system access.
After you switch

Tested again at 30 days

Models change. If quality slips, you roll back.

What we never ask for

Your keys or your systems

We work from a redacted usage export. No passwords, no API keys, no access.

Who it's for

Built for £1.5k+ a month of AI spend

Best fit when model, API or agent spend is serious, ideally £3k+. Below that, payback is usually too slow.

Any provider, tested the same way

The models

We test whichever models you run, or want to run, against each other on the same examples and the same pass mark.

Anthropic ClaudeOpenAI GPTGoogle GeminiMistralDeepSeekMeta Llama+ others

Names are shown only to say which models can be tested. ProofCut is independent and is not affiliated with, endorsed by or a partner of any of these companies.

Pricing

Quoted in writing

Written note

Free

Likely cause, rough save range if your models and volumes are clear, and the price.

Written diagnostic

£250

Deeper read, cause and test plan. Credited if you continue to a full engagement.

Full engagement

£750–£1,500

Eval harness, switch support, 30-day verification. A one-off project: fine if you leave after.

Not financial advice. No savings guaranteed until an eval passes.

Start with six questions.

Free
Request a free written note →

Written reply within 2 working days · no call · no keys