The AI bill checklist
Nine questions to ask before you renew anything. Every one of them has cost a
real company real money this month.
Maintained from a live daily log of open-model pricing. Last checked
today against the live board.
The questions
- What are you paying per million tokens — input and output, separately?
They are not the same number and the gap is 5x to 25x. A vendor quoting a single
"AI cost" is quoting the cheaper half.
- Is your bill set on list price, or on the price you actually signed?
Open-model prices move most days. On 14 Oct alone we logged 40 changes across 462
models — 16 rises and 15 cuts. If your bill is set on last quarter's numbers, it is wrong
in both directions and nobody is checking.
- Which models are you on, and are they still the cheap ones for your job?
On 13 Oct openai/gpt-oss-120b — a default for a lot of agent stacks — tripled from
$0.037 to $0.15 per million input tokens. On 14 Oct it fully reverted to $0.037. A 305%
hike undone inside 24 hours.
Source: our own change log, ai-price-watch, 13–14 Oct 2026.
- Where are your output tokens going?
Output is where the invoice actually lives. On 14 Oct DeepSeek Pro Latest output went
$0.7484 to $8.00 per million — a 969% move. On the same day DeepSeek V4 Flash 0423 output
fell 86%. Same market, opposite directions.
Source: our change log, 14 Oct 2026.
- Do you have a batch or off-peak path switched on?
Many vendors sell the same model far cheaper in batch. gpt-oss-120b batch is
$0.0296/$0.136 per million against $0.037/$0.17 standard. If your workload is not latency
sensitive, you are paying a premium for nothing.
Source: OpenRouter catalogue, 14 Oct 2026.
- Who inside the company can actually see the AI bill?
57% of finance leaders say they are not confident they are paying a fair price for
AI. Most of them cannot see the line item.
Source: Business Wire, 28 Sep 2026.
- If your agents spend money on their own, who approves each call?
Agent spend is now a funded problem: Fragment raised $12M this month specifically to
stop AI agents going rogue with company money. If yours can call a paid API unattended,
that is an unbudgeted line.
Source: Fortune, 14 Oct 2026.
- Can you run the sensitive half on your own hardware?
The Dutch tax authority cancelled its Microsoft 365 cloud rollout and moved to open
source instead. Cloud sovereignty is now a procurement question, not an ideology.
Source: ZDNet, reported 14 Oct 2026.
- What is the renewal date, and what has changed since you signed?
This is the whole checklist in one question. Prices, models and your own usage have
all moved since the contract was written.
Want the numbers run for your own bill? I audit a company's AI and software spend,
cut what can be cut, and take 30% of what I save. You keep 70%. If I find nothing, you pay
nothing.
Send me your bill — free read