Blog
What we learn from measuring AI spend at companies already running it in production.
-
What each customer costs you in AI
If you sell a product with AI inside, cost per customer decides your margin and your price. How to calculate it, why the average lies and what to do with expensive customers.
Read → -
Volume discounts and reserved capacity: when they turn expensive
Committing volume for a better price looks free, but it locks in your consumption exactly when optimization should be cutting it. How to decide with numbers.
Read → -
Caps, quotas and alerts: how to avoid the surprise bill
Almost nobody spots an AI spend spike until the invoice arrives. How to set limits per team and per function, and what to alert on before it is too late.
Read → -
What you pay for twice: retries, timeouts and discarded answers
Failed work is billed the same as good work. Retries, timeouts and outputs that fail validation can be a silent share of your AI bill.
Read → -
Distillation: when training your own small model pays off
Using the large model to teach a small one your specific task. How much the cost per request drops, what it costs to set up, and the volume at which it makes sense.
Read → -
How to move to a cheaper model without losing quality
The biggest saving lever is switching to a cheaper model, and nobody dares because they cannot measure whether it gets worse. How to build a minimal eval in a day.
Read → -
Why an agent costs far more than it looks
An agent does not make one request, it runs a loop. The history grows every turn and the cost with it. How to measure it per completed task and where to rein it in.
Read → -
Input tokens and output tokens do not cost the same
Output costs three to five times more than input on almost every model. What that means for your bill and how to trim the expensive half.
Read → -
Token pricing: the 2026 changes that raise your bill without you touching anything
Promotions with an end date, long-context tiers, data residency uplifts and a price rise already announced for January 2027. What changed at OpenAI, Anthropic and Google, and what to do about it.
Read → -
Own model vs API: when self-hosting pays off
When hosting your own model beats paying per API request: the real volume threshold, the costs nobody adds up, and how to decide with numbers.
Read → -
Why long context drives up your AI bill
Every context token is billed on every request. Why long windows multiply the bill and how to trim them without losing quality.
Read → -
How to monitor AI spend: which metrics to track
What to measure to control AI cost: cost per task, per user and per endpoint. Without those three metrics you do not know where you are overpaying.
Read → -
What an AI token is and how it is billed
A token is the unit AI models read and bill text in: roughly 4 characters or 0.75 words. How they are counted and why they decide your bill.
Read → -
Which model to use for each task: the biggest cost lever
How to pick the AI model by task: most are solved by a small model at a fraction of the price. Task to model to cost table and how to route.
Read → -
Batch API: half the cost if the task can wait
What the batch API is and how much it saves: around 50% versus real-time requests, in exchange for not being immediate. Which tasks fit and which do not.
Read → -
Prompt caching: pay once for the context that repeats
What prompt caching is and how much it saves: 50-90% on the context tokens you reuse between requests. When it pays off and how to turn it on without touching quality.
Read → -
How to cut your OpenAI bill: 7 levers that save the most
Seven ways to lower the cost of the AI API without losing quality, ranked by how much they save. Most take hours to apply, not weeks.
Read → -
GPT-4o vs Claude vs Gemini: API prices compared (2026)
How much each model costs per million tokens and, what really matters, how much each one ends up costing in your case depending on volume. The table and the why.
Read → -
RAG, fine-tuning or long context: which is cheapest for your case
The comparison is usually made on quality. But the decision almost always breaks on cost, and there the three options behave very differently.
Read → -
Cost per token vs cost per task: why your AI bill grows while prices fall
OpenAI cut API prices by up to 80% and plenty of companies watched their bill go up the following month. The explanation lies in the unit you are measuring.
Read →