Per-token AI costs dropped 67% last year. Goldman Sachs expects global token consumption to hit 120 quadrillion per month by 2030, a 24x increase from today. So your AI bill should be falling, right?

It isn’t. The Australian Financial Review reported this year that organisations are blowing through their annual AI budget allocations well ahead of schedule. The per-unit price falls, the total spend climbs. That gap is the core story of where AI costs are heading, and most small business owners haven’t been told any of it.

The bill shock is real and documented. It’s already hitting businesses, and understanding the math helps you control costs.

The token math nobody explains

AI charges by the token. About four characters each, roughly three-quarters of a word. Every prompt in, every response out, every document your AI reads before it answers: all tokens, all billed.

Analysis from Singapore-based Ai.cc covering 2.4 billion enterprise API calls found the blended cost across businesses sits at US$6.07 per million tokens, down 67% year-on-year. Buried in that number is an 8x price gap that explains a lot. Companies routing tasks to the cheapest capable model paid a median of US$2.31 per million tokens. Companies defaulting to frontier models paid US$18.40. Same output quality in most cases, eight times the cost. The difference is whether someone made a deliberate choice about which model to use.

Most businesses don’t make that choice. They pick a tool, use the top-tier model it defaults to, and pay frontier prices for tasks a cheaper model would handle just fine.

Why your bill is climbing

First, usage is spreading. What started as one person in the business experimenting has turned into a team, then a whole company workflow. More seats, more queries, more tokens. Silicon Valley’s AI bill is landing in Australian small business inboxes now, not because the tools got expensive, but because the usage got serious.

Second, the models themselves are using more tokens per answer. Modern reasoning models, the ones that “show their working” before responding, run extended thinking chains before producing output. That internal monologue costs tokens too. A question that cost 500 tokens six months ago might cost 2,000 tokens today if you’ve upgraded to a reasoning model. The answer might be better. The bill is higher regardless.

Third, most defaults point to the frontier. When you sign up to an AI tool, it puts you on the most capable model. Vendors have no incentive to offer cheaper models by default. From your perspective, there’s no reason to pay US$18.40 per million tokens for tasks that US$2.31 would handle.

The deeper issue: whoever controls your defaults controls your costs. The ASPI Strategist framed this as a national sovereignty question this week. For small business, the same logic applies one level down. Right now that’s your vendor setting the defaults. You can change that.

What to actually do about it

Start with an audit. Pull your AI invoices for the last three months and find out what you’re paying per tool. Some platforms break out usage by model; others just give you a total. If you can’t see per-model or per-task costs, that’s the first problem to fix. The 2026 AI cost roundup for Australian small business has a breakdown of which platforms give you visibility.

Then sort your tasks. Long-form writing, complex analysis, specific judgment calls justify a frontier model. Summarising a document, drafting a quick email, answering a FAQ: a cheaper model handles these just fine. Most AI platforms let you select the model per conversation or per task. Use that.

Treat AI spend like any other budget line. Set a monthly limit. Track it. Review it quarterly. The difference is discipline. Businesses that track spend hit US$2.31 per million tokens. The ones that don’t? US$18.40, and they’re wondering why their bills keep climbing when per-token prices keep falling.

📖 More deep dives: News Deep Dives

📊 Compare AI tools side by side | 💼 Free resources & AI prompt packs

: News Deep Dives →

Sources