What GPT-6 Luna means for AI agency chatbot costs
GPT-6 Luna runs a client chat conversation for about half a cent. Here is the math, what the model can't do (audio, for one), and why you shouldn't cut your monthly because of it.
A text chatbot for a local service business now costs a fraction of a cent per conversation to run on GPT-6 Luna, OpenAI's cheapest current model. OpenAI's model page lists it at $0.10 per million input tokens and $0.50 per million output tokens. For an agency, that means the model bill on a chat or SMS line is close to noise, and the number you should be pricing around is your own time. Voice is a different story, and I'll get to that.
Owners keep asking me whether a model that cheap is any good, so we read the model page and the Vapi changelog this week. The short answer is yes for scripted front-desk work, and no for anything that needs audio or a long chain of tool calls without care.
What did OpenAI actually release?
GPT-6 Luna is the small end of the GPT-6 family. OpenAI describes it as "our most efficient model for focused, high-volume tasks." Eesel's pricing write-up puts the launch on September 23, 2026, and I could not confirm that date on an OpenAI page, so treat it as secondhand.
I could confirm the rest on OpenAI's own page. Input is $0.10, cached input $0.01 and output $0.50, all per million tokens. The context window is 1,050,000 tokens, with 128,000 tokens of output. Text and images go in, text comes out, and audio is not supported. Batch and Flex tiers run at half the standard rate and fast mode at double. Function calling, structured outputs and streaming all work.
Vapi added it the week of October 5. Its changelog says gpt-6-luna is available for assistants in the API and dashboard, with reasoning support and a fast service tier. So if you build on Vapi, you can point an existing assistant at it today.
What does a chat conversation cost on it?
Vendor pages won't do this arithmetic for you.
Say a website chat runs ten customer messages. Your system prompt plus the business's services, hours and price bands is 3,000 tokens, and it gets sent on every turn along with the growing history. Each customer message is about 60 tokens and each reply about 120. Add it up across ten turns and you send roughly 38,100 input tokens and get back 1,200 output tokens.
At Luna's standard rates that is about $0.0038 for input and $0.0006 for output. Call it 0.44 cents. A thousand conversations come to about $4.41.
Those are my assumptions, not numbers from a client account. Swap in your own prompt length. Prompt caching would push the bill lower still. But even if I'm off by a factor of five, a client doing 300 chats a month is paying you for a model that costs them roughly a dollar.
Compare that to the platform you sell it on. Lumina's plan includes 500 chat messages a month, and our wholesale tier for a website plus AI chatbot is $550 setup and $125 a month. The model is not where that money goes.
Why isn't the model bill the interesting part?
Nobody ever lost a client over a four-dollar token bill.
Clients leave when the bot quotes a price that doesn't exist, double-books a Tuesday, or says something strange to a woman whose basement is flooding. A cheaper model does not fix any of that. It also doesn't fix the other cost center, which is you. We went through the per-minute math on voice in what a voice AI agent costs per minute, and the conclusion there carries over: raw usage is the small line. Hours spent babysitting prompts are the big one.
There is one real change, though. When the model costs almost nothing, you can afford to run more of it per client. A second pass that checks the reply against the price list before it sends. A 6 a.m. text to the owner summarizing last night's conversations. Each of those was a line item you hesitated over at higher rates. At $0.10 in, you stop hesitating.
Where does Luna fall short for an agency?
Read this part before you switch anything.
It does not do audio. The model page lists text and image input and text output only. A phone line still needs a voice stack, and speech-to-speech models are priced on a different scale. OpenAI's own gpt-realtime-2.1 announcement lists audio input at $32 and audio output at $64 per million for the full model, and $10 and $20 for the mini. So the cheap model helps your chat and SMS, and your text-based reasoning behind a voice agent if you use a chained setup. It does not make a phone call cost a cent.
Tool calling has a catch. Per the model page, Chat Completions supports function calling only when reasoning is set to none. If your booking flow depends on the model deciding when to call a calendar function, you either run it with reasoning off or move to the Responses API. Test that before you promise a client anything. On the realtime side, a commenter in that same OpenAI thread reported that the mini model stopped calling function tools in a SIP flow where the older mini worked, which is a good reminder that a cheaper model in the same family is not a drop-in swap.
Small models may drift on long instructions. I have no benchmark to cite, so read this as a prediction and not a finding: small models usually hold a short, specific prompt better than a 3,000-token rulebook. Test yours before you believe me.
Some clients shouldn't bother. If your client is a law office or a clinic where one wrong sentence is a liability, the model cost difference is irrelevant. Pay for the strongest model you can and put a human check on anything that touches advice.
How should I use this when I'm pricing a client?
Don't cut your price because the model got cheaper. Your client isn't buying tokens. They are buying the missed call that turned into a water-damage job worth thousands.
I would quote on the outcome, hold your monthly where it is, and spend the savings on reliability. Add the checking pass and the owner summary, and log every conversation so you can read the bad ones on Friday. If you want a way to frame that conversation, how much to charge for an AI buildout walks through it.
And try it before you trust it. Run Luna against your own demo transcripts, or against the trade you sell into most, and read twenty of the replies. The live demo is a decent place to see what a front-desk agent should sound like before you point a cheaper model at it.
FAQ
Is GPT-6 Luna good enough for a small business chatbot?
For answering hours, services, price ranges and booking requests, yes, as long as the prompt is short and specific. It supports function calling and structured outputs, so booking is possible. Read a sample of real conversations in the first week and tighten the prompt where it wanders.
Can I use Luna for an AI phone receptionist?
Not by itself. It handles text and images only. You can use it as the language model in a chained voice stack, with separate speech-to-text and text-to-speech, but a speech-to-speech model is a separate product with separate pricing. Vapi lists Luna as an available model for assistants.
Should I drop my monthly price because model costs fell?
No. Clients pay for booked jobs and answered calls, and the model is a rounding error in what they cost to serve. Spend the savings on checking passes and monitoring instead.
What does Luna cost compared to the realtime voice models?
Luna is $0.10 and $0.50 per million tokens for input and output. The full realtime-2.1 model lists audio at $32 in and $64 out per million, per OpenAI's announcement. Those are different kinds of tokens, so you can't read them as a clean multiple, but the gap shows why chat and voice get priced so differently.
What to do next
Pick one client with a website chat or an SMS line and move it onto Luna for a week. Read every transcript. If it holds, the margin on your text products just got a little wider, and you have a reason to add the extras above. If you don't have a chat client yet, start with the Lumina Sales demo and build the first one this week.