Cut your AI bill.
We find where your product overpays for AI, fix it in your code, and measure the saving on your own traffic. You pay only after you have seen the saving.
Pinned to Haiku and Sonnet. Running on Opus.
The code pins both helpers to cheaper models. On n8n Cloud's model proxy there is no helper API key, so both fell back to the run's main model.
84,000 tokens at full price, on every step.
A timestamp accurate to the second sat ahead of the static prompt, so no two requests ever matched and every call paid full input price.
Every follow-up paid for the whole conversation again.
The history replayed from the database was not byte-identical to what was sent the turn before, so every follow-up wrote the conversation to the cache again.
The prompt said ten panels. It sent 2, or 55.
The selection kept the wrong end of each list. A missing backslash also joined panels with a literal n instead of a new line.
A timestamp kept the prompt out of the cache.
The current time and the conversation context sat ahead of about 1,411 tokens of fixed rules, so nothing was cached.
The AI meter billed some requests at twice the real price.
Providers start the long-context rate elsewhere: 272K for 13 OpenAI models, 100K for Claude Haiku 5.5.
- No saving, no fee
- You approve each fix before we build it
- Measured on your own traffic
How it works
One fix at a time, each one measured.
-
We read your AI code
Under an NDA, we read the code that calls your models. In about two days you know how many things we found and how much they could save in total.
Free
-
We fix one thing at a time
Each fix starts with one page you sign: what changes, the saving we expect and the test it must pass. Then we build it. It goes live only when it passes.
Nothing is built without your signature
-
You pay from the measured saving
We compare your cost before and after on your real traffic, and you sign the number. For six months you keep half of that saving and we get half. After that, all of it is yours.
No saving, no fee
What we find
The waste is rarely one big mistake. It's a handful of small ones, multiplied by every request you serve.
- Prompts that never hit the cache
- A timestamp, an ID or a user name sits ahead of the fixed part of your prompt. No two requests match, so every call pays full input price. Found in PostHog Chatwoot Metabase
- The same context, sent again and again
- Documents, schemas or instructions are attached to every message in a thread, when once per conversation would do.
- Conversations that only grow
- History is never summarised or trimmed, so each turn costs more than the one before.
- Large models on small jobs
- A frontier model classifies, extracts or routes where a small one gives the same output for a fraction of the price. Found in n8n
- Retries you pay for
- Failed or timed-out calls are sent again with the full context, and each attempt is billed.
Six findings you can check
We audited open-source products and sent what we found to their maintainers. Every number is measured, and every case links to the public pull request or issue. Four lower the bill; one fixes what the model is told, and one what customers are charged.
-
n8nsavingPinned to Haiku and Sonnet. Running on Opus.
−62%cost per verification
#43441 -
PostHogsaving84,000 tokens at full price, on every step.
−92%cost per repeat planner step
#132240 -
MetabasesavingEvery follow-up paid for the whole conversation again.
21–64%saved per follow-up turn
#86814 -
ChatwootsavingA timestamp kept the prompt out of the cache.
−61%input cost, first call of a reply
#16604 -
GrafanacorrectnessThe prompt said ten panels. It sent 2, or 55.
55%of 1,000 dashboards sent the wrong panels
#137521 -
Twentybilling accuracyThe AI meter billed some requests at twice the real price.
~2×debited against the provider's price
#30141
Pricing
One fee per fix, paid from the saving it makes. No saving, no fee.
- Audit
- Free. We tell you what we found and what it could save.
- Each fix
- 3 times the monthly saving we measure. One-time, paid in six monthly instalments.
- No saving
- No fee. A smaller saving than we estimated means a smaller fee.
- Other fixes
- A fix that doesn't lower the bill has a price agreed before we start.
- Who it's for
- AI bills of about $5,000 a month or more.
- Open source
- Free.
How the saving is measured, and what happens if a fix breaks
Fair questions
Why would we give you access to our code?
We need to see only the code that calls your models, under an NDA, and you can remove the access any day. Anything we add, including the logging we measure with, goes through your own review. We did the same on PostHog, Metabase, n8n and others, where every finding is public.
How much of our team's time does it take?
One person who can sign. The first read takes about two days and needs only read access. After that, your engineers review and merge our pull requests like any others.
How is the saving measured?
Your cost per call or per conversation, over 7 to 14 days before the fix goes live against 7 to 14 days after, on your real traffic and at the same price list. You see the logs and you sign the number. Your growth doesn't raise our fee, and a provider's price cut isn't counted as our work. Each fix is compared with your cost just before it went live, earlier fixes included. You never pay twice for the same saving.
Will you change the model we use?
Only if you agree. Caching and pricing-tier fixes come first, because they don't change what the model answers. A model change needs a quality test on your own requests, with the pass mark agreed before we start.
What if a fix stops working?
We repair it, and that is included in the price. You never pay for a fix that isn't working.
Why not have our own team do this?
You can. These bugs sit in code that works, so nobody is looking for them, and looking for them is all we do. With us it costs nothing unless it saves.
Find out how much you could save.
Tell us roughly what you spend on AI each month. If an audit isn't worth your time, we'll say so.
Get a free audit [email protected]