On Sept. 22, the price of artificial intelligence dropped twice in one day. OpenAI cut the price of its GPT-6 Sol and Luna models in half compared with the previous versions, Matthias Bastian reported in the-decoder, and Anthropic released Opus 5.5 at a lower price than Opus 5. Ars Technica summed up the pair in its headline: "A little more for a lot less money."
For the companies building on these models, the news is simple: the same work now costs them less.
For a small business owner who pays for one AI employee, the useful question is a different one. Does a cheaper model change what you should be watching? Mostly, no. And the reasons are worth a few minutes.
The price of a word is not the price of a job
Look closely at how the cuts were described. Anthropic's own numbers, as Ars quoted them, put Opus 5.5's input and output tokens 20% below Opus 5. But the company says the savings on typical work are closer to 40%, because the new model also uses fewer tokens to finish a task.
That is the whole problem with watching a per-token price. Tokens are the small chunks of text a model reads and writes, and most AI providers charge businesses by the token. The price of one chunk tells you very little about the price of a finished job. A model that is cheaper per word but chattier can cost more per answer; one that is pricier per word but gets there faster can cost less.
For the companies building AI products, that math is their business. For an owner, it is a distraction. What you want to know is what the job cost and whether it got done.
It is easy to measure the wrong thing
The same trap exists without any tokens at all. It is tempting to judge an AI employee by how busy it looks: how many messages went through, how long it was "on." None of that says whether a customer got the right answer. Measuring how much the AI ran instead of what it got done is a mistake that is easy to make at any size.
The businesses that hired a second one
The third item answers a question that owners with one AI employee start asking around month two: would a second one help?
On Sept. 23, Mike Wheatley reported in SiliconANGLE that Ema, a company that builds what it calls "autonomous AI employees" for HR, IT and finance departments, raised 77 million dollars. The funding is the headline; one sentence further down is the part to keep. In the last year, he wrote, Ema's customers more than doubled their spending on AI workers on average, mostly by expanding the deployments they started with.
Read that carefully. The customers did not buy a bigger model or burn more tokens. They took something that worked in one place and put it to work in another one. "Our customers are not running experiments," Ema's chief executive, Surojit Chatterjee, said.
These are large companies, and the numbers are Ema's own. But the shape carries down to a shop with one front desk: growth came from adding a second worker where the first one had proved itself, not from making the first one busier.
What this looks like in a business with one AI employee
Put the three items together and the owner gets a simple picture.
The price of the model underneath is not your number. If you pay for an AI employee by the role, the drop in token prices doesn't reach your bill, and a jump wouldn't either. What reaches you is whether the work got done.
And "done" is countable without any token math. How many customers got an answer this month? How many of those answers were right, and how many needed you anyway?
Three checks an owner can run in one afternoon
1. Count what got done, not what ran. Take last month and write down three numbers: questions answered, questions you still had to handle yourself, and complaints about a wrong answer. That is the whole report, on one page.
2. Watch the ceiling, not the meter. If your AI employee has a monthly limit on replies, the useful signal is how close you get to it in a busy week. Getting close every month is the moment to think about a second one, not about squeezing the first.
3. Hire the second one for a second job. If the first employee is busy with one kind of work and another kind keeps waiting, the second hire is a different job, not more of the same one. If the same job has simply outgrown one employee, it's the same role again.
How we think about it
We build AI employees, so we are not neutral. Our answer to September's price cuts: an owner should be able to ignore them.
Each role on SmashOne is $49 a month, with one channel included; each next channel is $9. A role comes with 1,000 AI replies a month for text roles; the receptionist has 250 phone minutes. That is a hard cap, not a meter that keeps running: there is no overage and no automatic charge. A reply counts when it is generated.
If one role is not enough, you don't buy extra replies. You hire the same role again, and the quotas of that role add up. If you need a different job done, you hire a different role. Either way the bill is the number of employees on your team, not the number of tokens they used.
You can see this from your own chat, too: in the ChatGPT app you already use, you can ask how much of this month's quota is used, and start the next hire there in a few words, then confirm it once in your SmashOne account.
Two sample businesses, Tampa Pasta House and Casa Lista, run on SmashOne in public, so the work is visible before anything is bought.
Hire an employee at smashone.us.
The models will keep getting cheaper. The question worth asking is still the same one: what did your employee get done this month?
Sources
- the-decoder, Sept. 22, 2026, Matthias Bastian — "OpenAI's GPT-6 Sol and Luna cut prices in half but barely move the needle on performance"
- Ars Technica, Sept. 22, 2026 — "New Anthropic, OpenAI models make same promise: A little more for a lot less money"
- SiliconANGLE, Sept. 23, 2026, Mike Wheatley — "Ema raises 77 million dollars in funding to deploy AI employees across enterprise HR, IT and finance departments"
