OpenAI and Anthropic cut model prices within hours of each other

Two releases, one Tuesday

OpenAI shipped cheaper AI models within hours of Anthropic on Tuesday. That timing made a technical release feel less like a laboratory milestone and more like a price tag moving in public.

We’re used to AI news arriving dressed as spectacle, but the useful detail is quieter: routine work may become cheaper to test.

A café, clinic, or studio doesn’t need a research department to understand that pressure. It needs a narrow task. And it needs a boundary.

Cheaper mistakes travel fast.

OpenAI’s GPT-6 Sol and GPT-6 Luna meet Anthropic’s Opus 5.5

On Tuesday, September 22, CNBC reported that OpenAI and Anthropic released lower-cost artificial intelligence models within hours of each other. CNBC said OpenAI shipped GPT-6 Sol and GPT-6 Luna, while CNBC’s coverage described both companies as facing competition from open-weight rivals.

Engadget reported that Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Astra offered similar performance at lower costs than earlier releases. Engadget’s report also said Anthropic claimed Opus 5.5 scored better than GPT-6 Astra on agentic coding benchmarks.

Tech.co reported that the releases were claimed to be around 40% to 50% cheaper than previous versions. NDTV reported that OpenAI and Anthropic released their models days after calls to slow AI development, with both companies aiming at more business users.

The products matter. The price movement matters more.

The cost cut matters more than the launch theatre

We think the part to ignore is the race between model names.

A new name doesn’t answer the question a small team actually has: can this perform one useful task cheaply enough, reliably enough, and with enough oversight that it earns its place? The answer isn’t found in a benchmark, a launch graphic, or a stream of claims about general intelligence.

It is found in the gap between a customer’s question and a staff member’s reply.

Lower cost changes that calculation because a contained trial no longer has to carry the expectations of a grand transformation. You can use an AI assistant to draft replies, sort incoming questions, or prepare a first response without asking it to run the whole relationship. That is a smaller bet, and smaller bets are easier to stop when they go wrong.

But lower cost does not mean lower risk. A poor answer on a booking page can still lose a booking. A wrongly timed WhatsApp reply can still feel careless. An inaccurate Google listing response can still make a local business look absent.

We do not know whether these cheaper models will be dependable on the customer-facing tasks that matter to you.

That unknown is the point. Don’t buy the storyline. Test the behaviour.

The price war beneath the safety warnings

The industry talks in model releases. Your customers experience it in response times and clarity.

The floor is where it counts.

A lower-priced model makes a human-reviewed first draft more practical. It does not make an unattended customer conversation safe.

So the move is not to hand over the keys. It is to choose a small door, watch what comes through it, and keep someone near the lock.

A bounded test on your WhatsApp or booking page

You probably won’t do this because it sounds like another system to set up, another tab to check, and another source of awkward mistakes. That resistance is sensible. The answer is not a large rollout. It is a short test on a surface where you already know what customers ask and where a staff member can still catch a bad reply.

For AI costs for small business, the useful question this week is whether a cheaper model lets you test one customer-facing task without turning it into a new full-time job.

• Start with your booking page. Collect the questions that repeatedly arrive before someone books: opening times, availability, location, what to bring, or what happens after payment. Have an AI assistant draft answers using only the wording you approve, then make a staff member review every answer before it goes live. If you skip that review step, the booking page can become a polished source of wrong information. If the test creates more corrections than saved typing, stop it. For a broader view of where a small automation can genuinely save time, read How Automation Can Save You 13 Hours Every Week.

• Use your WhatsApp as a draft desk, not an autopilot. Set up an AI assistant to turn a new enquiry into a proposed reply that includes the customer’s question, the approved answer, and a clear handoff when the message needs a person. Your staff still presses send. If you skip the handoff rule, customers with unusual requests can get trapped in a conversation that sounds helpful but resolves nothing. Keep the trial narrow, and review the drafts on your Friday afternoon before deciding whether it belongs in the week ahead. How Smart Systems Can Transform Your Business and Free Up Your Time is useful when the aim is to remove repeat work without removing judgment.

Don’t begin with sales promises. Begin with the messages that make staff sigh because they have typed the same answer again.

And don’t measure the test by whether the AI sounds impressive. Measure whether the customer gets a clear answer and whether your team has fewer repetitive replies to write.

The next round of model price cuts

Watch for one signal: whether the cost of a tightly bounded, human-reviewed workflow falls enough that you can keep it running without cutting corners.

That signal will appear in the work itself. Look at your booking page replies, your WhatsApp drafts, and the questions that still require a person after the assistant has had a first pass.

If the assistant reduces repeat writing while staff corrections stay manageable, the price movement mattered. If corrections pile up, it didn’t.

The next meaningful release won’t be the one with the loudest name. It will be the one that makes supervised customer communication cheap enough to keep.