AI Model Prices Just Halved: Time to Recalculate Your Cost per Task

OpenAI has launched GPT-6 Sol and GPT-6 Luna at roughly half the API price of their predecessors, according to an IT Brief UK report dated 24 September 2026. For owners who pay for AI tools by usage or by plan, this is a good moment to recalculate cost per task and to ask the provider a plain question: does that cut reach my bill.

One caveat up front. Every figure below comes from a single technology news report. We could not open OpenAI's official page while writing, so check current prices before making a decision.

The reported numbers

According to IT Brief UK, prices per million tokens are as follows. A token is the unit of text used to measure model usage; input is the text you send, output is the text the model produces.

The comparison price for Sol is the promotional price of GPT-5.6 Sol. Other points from the same report:

• Sol and Luna sit below GPT-6 Astra, OpenAI's top model.
• Cached input token reads get a 90% discount. OpenAI links the lower prices to changes in caching and inference.
• Sol and Luna are available to Plus, Pro, Business, Enterprise and Edu users. Free and Go users get Luna only, in the desktop app.
• By OpenAI's internal evaluation, Sol makes about half the errors of its predecessor. That is OpenAI's own claim, not an independent test.

These are API prices in US dollars, the prices paid by developers and tool providers. They are not app subscription prices, and they are not necessarily what you pay.

Working out cost per task

The most useful way to read a price per million tokens is to turn it into a cost per task. The following is our own illustration with a simple assumption: one customer reply draft uses 2,000 input tokens (conversation history and instructions) and 300 output tokens (the reply).

Two lessons from that table. First, the gap between the two models is far larger than either price cut: in this illustration Sol costs about twenty times Luna for the same task. Second, at small business scale, the model cost of a light task is already tiny. The AI cost small business owners actually pay is often set less by token prices than by which model is used for which task, and by how the provider packages its pricing.

Token counts for your tasks may differ a lot from this assumption. Long documents, attachments and thick conversation histories push input tokens up.

Questions for your AI tool provider

This section and the rest are our advice. Most small businesses do not buy tokens directly. They use a tool with an AI model inside it, so the questions go to that tool's provider.

1. Which model is used for each kind of task? A short reply draft rarely needs a top-tier model.
2. Can I choose the model, or does the provider decide?
3. When upstream model prices fall, does my rate or quota change? When?
4. What is the billing unit: per message, per action, per user or per plan? What happens when the quota runs out?
5. Is there a usage report per task or per staff member, so I can see where the cost goes?
6. What data is sent to the model, and is it stored?

A good provider can answer all six without fuss. "It depends" is an acceptable answer as long as it comes with an explanation of what it depends on.

Do not treat model price as a provider's whole cost either. They carry development, support and storage. The point of asking is not to demand a discount but to understand what you are paying for.

Tasks now worth trying

When cost per task falls, the list of tasks that make sense to automate grows. But cheap is not the same as correct. The half-the-errors claim comes from the model's maker, and half the errors still means errors. The safe pattern: a cheap model prepares, a person approves.

Tasks that suit this pattern:

• Reply drafts for recurring questions, checked by staff before sending.
• Summaries of long conversations when a customer is handed to another staff member.
• Sorting incoming messages: orders, complaints, price questions.
• Drafts of payment reminders or quotation follow-ups.
• Turning meeting notes into a task list.

Tasks that should stay fully in human hands: special pricing decisions, replies to serious complaints, anything touching law, tax or health, and promises that bind the business.

How to recalculate in one hour

1. Write down the five tasks your team does most often through chat or documents.
2. For each, note how many times a week and how many minutes each time.
3. Ask your tool provider what the AI cost is for that kind of task, or estimate it from the usage report.
4. Compare that cost with the staff time saved, then subtract the time spent checking the output. Checking is part of the cost.
5. Pick one task, run it for two weeks, and record how many drafts were used unchanged, lightly edited or discarded.

That last number is more honest than any benchmark. If most drafts are discarded, the task is not a fit yet, however cheap the model.

Where Polynusa fits

Polynusa provides AI reply drafts that staff approve before they are sent, which is the prepare-and-approve pattern described above. Plan details are on the pricing page.

Lower model prices are good news, but the benefit only shows up when you know which tasks use AI, what they cost, and who checks the result.

Sources

• IT Brief UK, 24 September 2026: OpenAI launches GPT-6 Sol and Luna at lower API prices

FAQ

What are the API prices of GPT-6 Sol and Luna?

According to IT Brief UK on 24 September 2026, Sol costs US$2 per million input tokens and US$10 per million output tokens, and Luna US$0.10 and US$0.50. These are API prices, not subscription prices.

Does a cheaper model automatically lower my AI tool bill?

Not automatically. The tool provider decides which model it uses and how it charges, so ask directly whether the price cut is passed on to customers.

Which tasks suit a cheap model?

Repetitive tasks whose output a person can check easily, such as reply drafts, conversation summaries and message sorting, with staff approval before anything reaches a customer.