OpenAI’s new GPT-6.1 Sol costs a fifth of what its flagship model costs, and OpenAI says it gets close to the same results on the work that matters to most businesses. That is the headline from DevDay on 29 September, and it is a bigger deal for a small firm’s bills than for anyone’s benchmark table.
Here is what was actually announced, what we would tell a client about it, and the two or three things to check before you change anything.
The number: a fifth of the price
GPT-6.1 Sol is priced at $2 per million input tokens and $10 per million output tokens. OpenAI’s top model, GPT-6 Astra, costs $10 and $50 for the same. That is one fifth on both sides. Cached input, the part of a prompt the model has already seen, drops to $0.10 per million tokens, which is 95% below the standard input price.
Tokens are the unit AI is billed in, roughly three quarters of a word each. To make that concrete, take an illustrative job that reads 1 million tokens and writes 200,000 tokens back, such as summarising a large pile of supplier emails and drafting replies. On Sol that is $2 plus $2, so $4. On Astra it is $10 plus $10, so $20. Those are published tool prices, not a promise of what you will save, because your real usage will differ.
OpenAI says Sol nearly matches Astra on agentic coding, computer use and professional work. It is available now to Plus, Pro, Business, Enterprise and Edu users in ChatGPT and Codex, and to developers through the API as gpt-6.1-sol. The word to hold on to is “nearly”. A near-match on OpenAI’s chosen tests is not a guarantee on your invoices, your customer emails or your spreadsheets.
What we would tell a client about it
The same lesson keeps coming back every time a vendor cuts a price: do not move everything, move the routine work. We use the top model for the thinking, the planning and the awkward edge cases, and a cheaper tier for the repetitive jobs where a slightly weaker answer costs nothing. Sol at a fifth of the price makes that split easier to justify, because the cheaper tier is now close to the top one for many tasks.
The practical test takes an afternoon. Pick one repeat job, such as sorting enquiries, drafting quote follow-ups or tidying a customer list. Run twenty real examples through both models. Count how many answers you would send without editing. If the cheaper model gets 18 of 20 and the expensive one gets 19, you have your answer. If it gets 12, you do not.
The trade-off is worth naming. A cheaper model that needs more retries can eat its own saving, and a cheaper model that makes a confident mistake in a customer email costs more than the tokens ever did. Price per token is not the same as price per finished job.
The other launch: dots, always-on agents
The second announcement was dots. OpenAI describes them as always-on agents that run on GPT-6 Astra, each with its own cloud computer and access to more than 4,000 apps. You reach a dot by text message, Slack, Teams or a phone call, and it works on tasks such as keeping track of a project, preparing for an appointment or following up a to-do list, coming back to you only when something needs your review.
That is a genuinely different idea from a chatbot you open when you need it. It is closer to a junior assistant who never logs off. For a small business that is attractive, and it is also exactly why we would slow down.
The price and the access come first. Your first dot is included with ChatGPT Pro ($100, $200 or $500 a month) or a Business Premium seat, which is listed at $125 per user per month, or $100 per user per month billed annually. Reports on the launch say Pro access to dots excludes the UK, the EEA and Switzerland at first, while Business Premium works across supported regions. OpenAI has not published what extra dots will cost or what each one is allowed to do each month. Check the current position on OpenAI’s own pages before you plan around it, because availability is changing fast.
The uncomfortable part is access. A dot that can reach 4,000 apps can also reach your customer records, your accounts and your inbox. Under UK GDPR you remain responsible for that data whichever tool touches it, and “the agent did it” is not a defence. An always-on assistant makes mistakes at the same speed it does everything else, and it does them while you are not looking.
Three things to do this week
- Pick one routine job and test it on Sol. Use twenty real examples, judge the answers yourself, and write down the result. Do not rely on anyone else’s benchmark.
- Look at your current AI bill. If most of your spend goes on the top tier for routine tasks, a cheaper tier is the first place to look, whichever vendor you use. Anthropic and Google both run tiered models too, so compare like for like.
- Write down what an agent would be allowed to touch before you buy one. Start with read-only access to one app, keep sending, paying and deleting behind a human approval, and widen only when it has earned it.
What we are watching next
Price cuts in this market have come in waves for months, and each one changes which model is sensible for which job. The safest habit is to treat your model choice as something you review every quarter rather than something you set once. The second thing we are watching is whether UK access to dots widens, because an agent you cannot legally or practically use is a headline and not a tool.
If you want a second pair of eyes on which jobs in your business suit a cheaper model and which should stay with the best one, that is the kind of thing our free 30 minute AI audit covers.
Sources: The Next Web on GPT-6.1 Sol pricing, VentureBeat on GPT-6.1 Sol, OpenAI on ChatGPT Business premium seats, Fortune on dots.