Three years ago, running one serious question through a top model cost real money. GPT-4’s API launched in March 2023 at thirty dollars per million input tokens and sixty per million output. By August 2026, several capable models charge under a dollar for that same million tokens. That is not a sale. That is a different economy, and it quietly changes the cost of AI for small business owners who once assumed this technology was priced for enterprises.
The drop is real, and it is worth understanding before deciding what to do about it. It is also easy to misread. Cheaper does not mean easier, and the bill that fell is not the one that was ever really holding most small businesses back.
The numbers behind the fall
Late in July 2026, OpenAI cut its GPT-5.6 Luna model by roughly 80 percent and Terra by 20 percent, according to pricing trackers compiled by AI Magicx and Spheron. Google’s Gemini 3.6 Flash dropped from about nine dollars to four-and-a-half for a combined million input and output tokens. On the low end, DeepSeek’s V4-Flash now lists near fourteen cents per million input tokens.
Zoom out and the trend is steeper still. Independent comparisons from AI Magicx and aisuperior put the industry-wide decline at 80 to 90 percent over two years, and more than 90 percent since 2023. Prices in this category tend to fall every three to six months as providers optimize their infrastructure. For a small business, the practical takeaway is simple: whatever a task cost to automate last year, it very likely costs a fraction of that now.
What becomes worth doing
Falling prices do not just make the same work cheaper. They make new work possible. AI Magicx notes that personalizing content for a single user in real time was effectively off the table at ten cents or more per personalization. At a tenth of a cent, it is viable for a small online store, a course creator, or a local service business. The same shift applies to summarizing every support email, drafting first-pass replies, tagging incoming leads, or checking documents for errors. None of these were unthinkable before. They were just not worth the per-use cost until now.
This is also where the choice of model matters less than it used to. When capable models sit within cents of each other, the more useful question is not which one is cheapest but what is now worth automating that was not before. That reframing sits close to an earlier point on this site about how no single vendor owns a small business AI stack anymore, and how the recent jumps in model capability keep widening what a modest budget can reach.
The cost that didn’t drop
Here is the part the pricing charts leave out. As analyst Pasquale Pillitteri puts it in a 2026 breakdown, the list price hides the real one. Reasoning and agent-style models emit far more tokens per task, so a low headline rate can still add up. Batch processing and context caching cut rates by 50 to 90 percent, but each adds a layer of setup that a busy owner has to learn or pay someone to manage.
And the bigger cost was never the tokens. It is the time and attention it takes to fold these tools into daily work. Most people paying for AI today are not using it anywhere near its potential, and the reason is rarely money. It is the learning curve, which has compressed hard and is more tiring than it sounds. That gap, not the invoice, is what usually stalls adoption, a point worth sitting with alongside why AI adoption stalls on know-how rather than tools.
One small way to use the news
Cheaper AI is genuinely good news for a small business, and it can also be one more thing to feel behind on. Both are true. The way through is not a big platform decision. It is one task. Take a single repetitive job, run it through one of the cheap, capable models for a week, and actually measure what it saved in time and what it cost in tokens. The number will almost certainly be small. Starting at all still puts a business ahead of most.
The prices will keep falling. The harder question is quieter: once nearly everything is affordable to automate, what is actually worth a small team’s limited attention to build first?
