Why we built Burnguard
The AI bill nobody checked until the end of the month
Building on OpenAI or Anthropic is easy. Knowing what it's costing you, day to day, is surprisingly hard, and the first sign of a problem is often the invoice.
By the Lowtide team · · 5 min read
The problem we kept seeing
AI API spend doesn't behave like other software costs. It moves with usage, with prompt length and with which model a developer happened to pick. A retry loop, a background job that runs more often than intended, or a switch to a larger model can multiply a day's spend without anyone noticing.
It's a widespread problem: Futurum's September 2026 survey of 1,636 technology buyers found 46.9% of enterprises say their actual AI spend is above their planned AI budget.
There's a second, quieter risk. Providers retire models on fixed dates. OpenAI, for example, has published 11 December 2026 as the shutdown date for several GPT-5 and o3 snapshots. After that date, calls to a retired model simply fail.
What goes wrong today
- Spend lives in each provider's console, and someone has to remember to look.
- A monthly budget alert tells you a threshold was crossed, not that yesterday cost five times a normal day, or which model caused it.
- Teams on more than one provider have no single view.
- Retirement notices arrive as emails to whoever opened the account, often someone who's since moved on, and get missed until something breaks.
How Burnguard approaches it
Burnguard connects to OpenAI and Anthropic with read-only admin keys. Every morning it reads yesterday's costs, compares them with your normal pattern and projects the month from completed days, so a partly finished day doesn't distort the forecast.
You get an email when spend jumps, when the month is heading over your budget, or when a model you actually use has a retirement date coming up. Retirement warnings arrive at 90, 30 and 7 days, with the recommended replacement. The schedule comes from the providers' own deprecation pages and is refreshed weekly.
It can't run up costs or change anything in your account. The keys are read-only and encrypted.
Example · Agencies and consultancies
A digital agency re-billing AI costs
An agency runs AI features for six clients on one OpenAI organisation, normally spending around £40 a day. On a Tuesday, a client's chatbot starts retrying failed requests in a loop and the day costs £310.
Without daily checks, that runs for three weeks before the invoice. With Burnguard, the spike email arrives on Wednesday morning showing the day's cost against the 7-day average, the dashboard's model breakdown points to the client's chatbot, and the loop is fixed that day. The same week, Burnguard flags that two client projects still use a model snapshot due to be retired, with 90 days to switch.
This is an illustrative example, not a customer story.
What you get in the first month
- One view of OpenAI and Anthropic spend, by day and by model
- A month-end projection based on completed days
- Same-week warnings when spend jumps or a budget is at risk
- A list of the models you use that have retirement dates, with replacements
Who it's for, and who it isn't
A good fit
- Startups shipping AI features
- Agencies and consultancies running AI work for clients
- Finance or engineering leads who own an AI budget
Probably not for
- Teams using AI only through a chat subscription like ChatGPT Team, with no API usage
- Usage through Azure, Bedrock or Vertex AI, which is not yet supported
Sources
- Futurum: 46.9% of enterprises report AI spend over budget (September 2026) (opens in a new tab)
- OpenAI API deprecations (opens in a new tab)
- Anthropic model deprecations (opens in a new tab)
Examples in this article are illustrative, not customer stories. Nothing here is legal advice.