Cheaper AI Still Needs a Website: What the GPT-6 and Claude Price Cuts Mean in Dallas
On September 22, 2026, OpenAI shipped GPT-6 Sol and Luna at about half the API price of the 5.6 series, and said Sol makes about half as many factual mistakes as the model it replaces. The same day Anthropic shipped Claude Opus 5.5 at roughly 40 percent less to run than Opus 5, with flagship-level performance on the jobs Anthropic cares about. Dallas owners read that as “AI got cheaper, so we should add a chatbot.” The useful reading is narrower. The model is now cheap enough to sit inside the website or the app you already needed. It is not a substitute for that website.
Quick answer: The September 22 price cuts make an assistant inside a real product affordable. They do not make a rented chat tab a business. Wire the model to your services, your service area, your hours, and your lead form. Keep the page quotable for Google’s AI Overviews. Measure booked work, not how clever the bot sounds.
- GPT-6 Sol and Luna: about half the API price of the 5.6 series, with a lower claimed error rate.
- Claude Opus 5.5: about 40 percent less to run than Opus 5, released the same day.
- A Dallas roofer, clinic, or new headquarters still needs pages a person and an answer engine can cite.
- Put the model on the work the business already does: qualify a lead, draft a reply, book a time, explain a price range.
What actually shipped on September 22
OpenAI’s note on Sol is specific. On an internal factuality check built from real conversations where users flagged mistakes, Sol made about half as many mistakes as its predecessor, at Astra-level reliability and a lower price. Luna is the smaller one, rolling out to more people, including free tiers in the desktop app. Anthropic’s Opus 5.5 is the first release since both labs publicly talked about slowing the frontier. The price move is the part a local operator can use this quarter.
None of that ships with your menu, your ZIP codes, or your crew’s calendar. A general model will invent a service area if you ask it to sound helpful. The system that does not invent is the one that reads your pages and your CRM and refuses to answer past them.
Where the cheaper model belongs
Three jobs are worth the integration. Everything else can wait.
- Answer the question your best page already answers. Service, city, price range, and the next step, in the first sentences. That is what Google’s auto-expanded Overviews can cite, and it is what a model on your site should repeat without decorating.
- Qualify the lead before a person does. Trade, timeline, address, budget band. Write the result into the CRM you own. The first-party record is the point. The chat transcript is not.
- Draft the follow-up your team already sends. Estimate recap, appointment confirmation, “we are on the way.” A person still sends it. The model saves the blank page.
A chatbot in the corner that can talk about anything is the expensive version of this, even at half the token price. It burns trust the first time it promises a same-day roof in a city you do not serve.
What not to buy with the savings
Do not buy another login. Owners already have ChatGPT, Gemini, or Claude on a phone. Customers will not create an account to ask what you charge. They will ask Google, or they will land on the ad Google’s AI Max chose for them, and they will leave if the page does not answer.
Do not turn the model loose on your ads. The platform is already choosing more of the query and the creative. Your job is the page and the conversion, which you can check. Their job is the auction, which you cannot.
What a Dallas build looks like
Start with three pages: the service, the city, and the proof. Each one leads with a sentence a model can quote without lying. Then add one assisted flow. For a service business that is usually “tell us the job and the ZIP, we will say whether we cover it and how to book.” For a company that just opened an office, it is “what we do, who we hire, and how to talk to Dallas,” which is the first 30 days online for a new headquarters.
InterGlobal builds that as a website, an app, or both. The model is a part. The system is the pages, the form, and the record of who asked.
Want the model inside the product, not beside it?
We will map the three pages worth citing and the one flow worth automating, then build it on a site you own.
Book a call