For a while, the honest answer to “should my small business build on these tools yet” included a quiet worry about the bill. That calculation just changed. Through 2026, a price war has pulled AI costs for small business down to a level that was simply not on the table a year ago, and it is worth understanding what actually moved before rushing to spend the savings.
The shift is real. It is also easy to misread.
How far AI costs for small business have fallen
The headline number is steep. Large language model API prices fell roughly 80% between early 2025 and early 2026, according to industry tracking summarized by Tom’s Hardware. Measured a different way, the blended cost of running these models dropped about 67% year over year, from around $18.40 to $6.07 per million tokens between the first quarter of 2025 and the same period in 2026.
Much of the recent drop traces to a single day. On July 9, 2026, three leading labs released new models within hours of each other, and the competition on price accelerated from there.
What the new price tags look like
Concrete numbers help. OpenAI’s economy tier, GPT-5.6 Luna, runs at roughly $1 per million input tokens and $6 per million output tokens. Google prices Gemini 3.5 Flash at about $1.50 input and $9 output, with cached tokens as low as $0.15. Anthropic listed Claude Sonnet 5 at an introductory $2 and $10 through the end of August 2026. To translate that: processing 50 million input and 10 million output tokens in a month, a meaningful amount of work for a small operation, lands near $110 on an economy tier, per pricing analyses collected across the sector.
For a small business weighing whether to automate a recurring task, that is the difference between “interesting someday” and “worth trying this month.” Many of these tasks now sit inside tools a business already pays for, from industry-specific AI built for one job to agents already bundled into everyday software.
The catch nobody prints on the price sheet
Cheaper per token is not the same as a smaller bill. This is where the two sides sit together. As prices fall, usage tends to rise, because tasks that were not worth automating suddenly are, and agent-style tools can loop through many calls to finish one job. The unit cost drops while the total can climb. A falling sticker price also does not remove the harder work: choosing among a menu that now runs to well over a hundred models, wiring the tool into a real workflow, and checking its output. That is its own kind of overwhelm, and it is fair to name it.
So the more useful question is not “which model is cheapest.” It is “which repeatable task is now cheap enough to be worth automating, and can its output be checked easily?” Price makes the experiment affordable. It does not make the judgment for you.
One small step this week
Lower the bar. Pick a single recurring task that eats time, drafting first-pass replies, summarizing invoices, cleaning up a spreadsheet, and estimate what it would cost at the new prices. Most owners find the number is smaller than expected, often a few dollars a month rather than a subscription-sized commitment. Run it once by hand through an economy-tier model and see whether the output is good enough to check quickly. If it is, that is real progress. If it is not, that is useful information too, and it cost almost nothing to learn.
The price of trying just dropped to near zero. What has not changed is the part that always mattered: knowing which task is worth handing over, and staying close enough to catch it when the tool gets something wrong.
