Guides
Plain-English breakdowns of what it costs to run AI.
Spot vs on-demand GPUs: $4.00/h instead of $7.99/h, if your job survives interruption
Verda spot vs on-demand prices on 2026-10-18 for six GPUs, which jobs belong on spot, which do not, and why the logic mirrors the LLM Batch API.
Same GPU, up to 28× the price: the provider spread on 2026-10-10
One chip, every tracked provider, 1 GPU on-demand on 2026-10-10: the V100 spread is 28×, the B200 spread is 1.3×. Check the provider before the chip.
LLM API price changes, September 2026: who cut, who doubled
Output prices tracked daily, 2026-08-31 to 2026-10-01: DeepSeek V4 Pro fell 77% while GPT-5.6 Sol Pro and GLM 5.3 Flash doubled. Recheck your model monthly.
The Batch API: half price for anything that can wait
Every major LLM provider has an asynchronous batch lane at roughly 50% of the normal per-token price. If a job does not need an answer in seconds, it belongs there.
Prompt caching: the first lever to pull on a RAG bill
If your requests share a long opening block - a system prompt, a document, a tool schema - you are paying full price to recompute it every call. Here is how much caching saves and how to turn it on.
What does it really cost to run an LLM?
A full breakdown: how per-token pricing works, why the same model costs different amounts for a chatbot, a RAG app and an agent, and when a rented GPU beats the API.
Six levers to cut your LLM API bill
Before you switch providers, pull these levers - most teams leave 40-70% on the table.
Is it cheaper to self-host, or just use an API?
Renting a GPU looks cheap next to a per-token bill - until you count the idle hours. The arithmetic that decides it.