AI just got 50% cheaper. What can your company automate now?

When a new generation of AI models arrives, the conversation always starts the same way. Is it smarter? Does it code better? Does it make fewer mistakes?

This time there is another question, far more useful to anyone running a business: what does it cost to put it to work?

On September 23, 2026, OpenAI released two new models, GPT-6 Sol and GPT-6 Luna. Sol now costs $2 per million input tokens and $10 per million output tokens, against $4 and $20 for its predecessor. Luna dropped to $0.10 and $0.50. The company attributes the reduction to inference and caching improvements, and confirmed the new prices carry no expiry date.

And it was not an isolated move. The same day, roughly ninety minutes earlier, Anthropic released Opus 5.5. This is not one company running a promotion. It is competitive pressure, and that is what makes the trend relevant to anyone setting budgets.

It may look like a conversation for developers. It is not. When the price of intelligence falls, tasks that did not justify automation yesterday can start justifying it today.

The problem was never only whether AI could do it

Imagine a company receiving 200 sales enquiries a week. Somebody has to read each one, work out what the person wants, check whether information is missing, classify the opportunity, log the contact, prepare a reply and define the next step.

We have known for a while that AI can help with several of those. The business question is different: what does it cost to do this 200 times a week? Then 2,000. Then 20,000.

A demonstration can be impressive and still make no economic sense. Which is why cost per task matters more than a score on a test.

The figure that matters is not in the chatbot

OpenAI tested Sol on AutomationBench, which evaluates agents on complete processes using 47 tools across sales, marketing, operations, customer support, finance and HR.

In the published results, the company states that Sol at high effort outperforms a competing model at roughly 9% of that competitor's cost per task, and that Luna improves by 5.4 percentage points at a cost per task 58% lower than its predecessor.

Two caveats, and they matter. First, these benchmarks are OpenAI's own, with competitor figures taken from public reports. Second, a benchmark is a test environment. A real implementation adds integrations, infrastructure, databases, monitoring, development, maintenance and human validation.

What the test shows is not a price. It is a direction: the computational cost of doing useful work is falling quickly.

When a technology gets cheap, the question changes

While it is expensive, we use it only on the most valuable problems. When it gets cheap, we start using it on small ones. That is what happened with storage, with bandwidth and with cloud computing.

Imagine a four-minute administrative task. A person does it perfectly well, so nobody ever considered automating it. But it happens 300 times a month. That is 1,200 minutes, or 20 hours, every month.

This is where small tasks turn into large opportunities.

The opportunity for smaller companies is not in giant processes

When people talk about digital transformation, they picture replacing the ERP or rebuilding the whole sales process. AI opens another possibility: automating hundreds of small moments.

A company may not need a hundred-thousand-euro project. It may need to solve five things: classifying enquiries, extracting information from documents, preparing replies, updating systems and following up.

Take a concrete case. A construction firm receives a quote request with photos, a description, an address and perhaps a bill of quantities. Today somebody opens everything, reads it, organises it, looks up previous prices and prepares a first structure.

This does not mean a machine should decide the final price of a job. It means it can extract quantities, organise the work, identify what is missing, compare with previous jobs and prepare a first draft, leaving the economic decision to a person.

Saving 30 minutes sounds small. If the firm produces 100 quotes a month, that is 50 hours.

Choosing the model becomes strategy

There is a common mistake: always using the most powerful model available. It is the equivalent of putting your most expensive professional on administrative work.

Some problems justify the best model available: complex analysis, difficult decisions, advanced programming, ambiguous situations. But classifying a simple enquiry does not. Extracting a number from an invoice does not. Normalising a product name probably does not.

That is why OpenAI now positions different models for different levels of cost and complexity, with Astra at the top, Sol for complex work with better efficiency, and Luna for high volume at low cost.

The architecture of a good solution starts to look like a team. You do not put the most expensive person on everything. You pick the right resource for each job.

The mature question stops being "can we do this with AI?" and becomes "what is the cheapest model that does this to the quality we need?"

Cheaper is not automatically more profitable

There is a danger in falling prices. When something gets cheap, we use more of it, and we can end up automating things that should not be automated.

A company can spend little on models and a lot on complexity, maintenance, errors, fragile integrations and human time spent correcting output. The model cost is a small slice. A cheap automation producing wrong work is still expensive.

The calculation that matters

Few business owners need to know how many tokens an email contains. The metric is different: cost per useful result. What does it cost to qualify a lead, process an invoice, prepare a quote, resolve a support request, analyse a document or produce a report.

Run the numbers with your own figures. Imagine a process taking 12 minutes, happening 500 times a month, with an average internal cost of 15 euros an hour. That is 100 hours, around 1,500 euros of work a month.

If a solution cuts 70% of that work and costs 300 euros a month across model, infrastructure and maintenance, there is an interesting economic conversation. If it saves ten minutes a month, there is not.

A simple method, in six steps

1. Find volume. A two-hour task done once a year is not a priority. A three-minute task done ten thousand times might be.

2. Calculate the current cost. Hours times average cost. It does not need to be exact, it needs an order of magnitude.

3. Assess the risk. Sending a newsletter and authorising a payment do not carry the same risk, and should not carry the same autonomy.

4. Automate a part. You do not have to eliminate the whole process. Often it is enough for the machine to prepare and a person to validate.

5. Measure. Time before and after, cost before and after, errors before and after.

6. Only then scale. If it works 100 times, look at 1,000. That beats starting with "let's implement AI in the company".

The price of AI is falling. The price of disorganised processes is not.

This is the irony.

You can have the cheapest model in the world. If the data is scattered across five spreadsheets, three inboxes and two WhatsApp groups, the problem remains. If nobody knows how a particular decision gets made, the machine does not know either. If the customer system is out of date, it works from wrong information. If products have no attributes, it cannot invent the right ones.

Falling prices remove one barrier. They do not fix disorganised processes. We wrote about that difference in 2026 is the year AI stopped answering and started working, and it remains the main brake we find inside companies.

The advantage is no longer having access

Computing, storage and bandwidth all became cheap. The same is now happening to artificial intelligence, with several suppliers pushing in the same direction.

Which means a company's advantage will rarely be having access to AI, because everyone will. It will be knowing where to use it, which processes to connect, what information to supply, when to use a cheap model and when to use an expensive one, where to require human approval, and how to measure whether it pays.

And there is a change of metric worth more than the rest. For two years many companies measured AI adoption by the number of licences. That is a weak metric. Having thirty licences is not digital transformation. Having a process that went from 20 minutes to 5 might be.

The next revolution may be unspectacular

There will probably be no day when a company wakes up transformed.

It will be a ten-minute process becoming two. A quote that took an hour ready in fifteen minutes. A team that handled 100 requests handling 300. A catalogue that took weeks to review being audited daily. A Friday report appearing by itself.

Small things, many of them. And it is precisely when the cost of intelligence falls that they start to make sense.

The news this week is that new models launched. The news that matters to businesses is different: it is getting cheaper to turn artificial intelligence into work done. When the cost of a technology halves, it is worth reopening the list of processes you decided last year were not yet worth automating.

Some of them may already have a different answer.

Frequently asked questions

What do these models cost through the API? GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 and $0.50. These are prices announced by OpenAI on September 23, 2026 and may change.

Did AI really get 50% cheaper? The API prices of these two models fell by roughly half against the previous ones. That does not mean a business project becomes 50% cheaper, because the model is only one slice of the total cost.

What can a small company automate? Good candidates include classifying enquiries, extracting information from documents, preparing replies, triaging support, updating data and reviewing catalogues. Viability depends on volume, risk and the systems involved.

How do you calculate the return on an automation? Start with the current cost of the process: frequency times time per run times average cost of the work. Compare that with the human time that remains, plus model, integration and maintenance costs. The metric is cost per useful result, not cost per token.

Does a small company always need the most advanced model? No. The best approach is the least expensive model that performs the task to the quality required, reserving more capable models for complex problems.

How Digito helps

Before asking where artificial intelligence could be used, we ask a different question: where is your company spending hours on work that repeats?

We analyse processes, calculate the automation potential using your own numbers, and build agents and automations connected to the systems you already have. We handle the integrations and custom software that let this run without manual work in between, and we take on the high-volume cases, such as online store catalogues with thousands of products to review.

And we say when it does not pay off. If the analysis shows the process needs tidying first, or that the volume does not justify the investment, we say so rather than sell a project.

If you want to understand where AI could genuinely save time in your case, let's talk about it.

Sources