Voice AI agents can now act while the caller is still talking
OpenAI's Decisions API went generally available on October 22, and GPT-Live can now call it while a caller is mid-sentence. Here is what shipped this month, what it costs, and the four places it does not help you.
Lumina12 min read
Yes. As of October 2026, a voice agent can read a caller's half-finished sentence and start doing something about it before the caller stops talking. OpenAI's changelog entry for October 15 describes GPT-Live gaining native Decisions integration, "letting apps act while the user is speaking or react to Live output." A week later, on October 22, the Decisions API went generally available on gpt-6-luna. The two together are the most useful thing to happen to voice agents since full duplex, and almost nobody selling AI phone systems has updated their pitch for it.
This is a roundup of what actually moved this month, written for someone running a one to ten person AI agency. The mid-call action change is the headline. The rest is smaller but still worth twenty minutes of your Sunday.
What is the Decisions API, in plain terms?
It is a cheap, fast classifier endpoint. You hand it text, an image, or JSON, and a list of questions. It hands back typed answers. OpenAI's guide says it returns those answers "about 10x faster than the Responses API," and it runs only on gpt-6-luna, the small model released on September 22 at $0.10 per million input tokens and $0.50 per million output.
There are three question types, and the shapes matter because they map onto things a receptionist actually decides:
predicate, which returns an estimated probability from 0 to 1 that something is true. Is this an emergency. Is this caller an existing customer. Is this person asking about price.
choice, which returns one value from a set you supply. Which department. Which service. Which of your four appointment types.
score, which returns a weighted average across 2 to 10 ordered levels. How urgent, on a scale you define.
At general availability the per-request limit went from 200 questions to 400. You will never need 400 on a phone call. You might need twelve.
If you have read our piece on what a voice AI agent costs per minute, this is the part of the stack that was previously either slow or hand-rolled. Most agencies either waited for the main model to finish reasoning, or wrote a pile of keyword matching and called it triage.
What does acting mid-sentence actually look like on a call?
The first is delegation. You define choice questions in session.delegation when you open the session, and GPT-Live decides on its own when to ask them. With native delegation, Live calls Decisions directly and skips a round trip through your server. With custom delegation, Live pings your app and your app calls Decisions, which is slower but lets you see and log every step.
The second is transcript triggers. Entries in session.triggers fire on session.input_transcript.delta, which is the caller's speech arriving in pieces, or on session.output_transcript.delta, which is the agent's own speech. Because the input triggers fire on partial transcript, the system can react while the caller is still mid-sentence. The documented example is cosmetic, mapping "I got the job!" to a happy expression and a nod for an animated character. The useful version for our market is not cosmetic.
Picture a water damage restoration line at 11pm. The caller starts with "hi, so my basement is, uh, there's water coming up through the". A predicate question on the partial transcript crosses your emergency threshold before the sentence ends. By the time the agent finishes its acknowledgement, the dispatch lookup has already run, the on-call tech's availability is already in hand, and the agent says a time instead of saying "let me check on that for you."
That gap is the whole thing. The pause after a caller finishes talking is where AI phone systems get caught. A homeowner with water in the basement does not consciously time the silence, but they feel it.
Two mechanics worth knowing before you promise this to anyone. Results can arrive out of order, so the guide tells you to apply the newest relevant result per direction and match source_event_id to the delta that produced it. And questions can change mid-session with session.update, but you have to wait for session.updated before you add a decision item for that step. Get either wrong and you ship an agent that acts on a stale read of the conversation, which is worse than an agent that pauses.
Why this matters more for the trades than for a tech demo
Because the trades buy on the phone call, and nothing else.
The buyer side has already moved. The US Chamber of Commerce's 2026 Empowering Small Business report, published October 2, puts small business AI use at 66%, up 8 points from 2025 and almost three times the 23% share in 2023. Two thirds of your prospects already use AI for something. That changes your first call. You are no longer explaining what AI is. You are explaining why the thing they tried did not work on their phone line.
Usually it did not work because it was slow and it sounded like it was reading. The same report found 78% of AI-using small businesses say restricting access would slow their growth, and only 6% say AI is enabling headcount reductions against 47% who say it is creating jobs today. Read that second pair carefully, because it tells you how to pitch. Owners are not buying AI to fire the front desk. They are buying it because the front desk cannot answer at 11pm and they are not hiring a night shift.
An agent that acts while the caller talks is the first version of this technology that sounds like a person who is already looking something up. That is a demo difference, not a spec sheet difference, and demo differences close deals. Our demo call playbook covers how to set that moment up so the prospect notices it instead of talking over it.
What does it cost to run?
Cheaper than you would guess, and the voice layer is still the expensive part.
GPT-Live voice sessions are $0.05 a minute, billed per second with no rounding up, with backend model and tool usage billed on top. The Decisions calls on gpt-6-luna are a rounding error against that. A phone call's partial transcript is a few hundred tokens, and cached input on Luna runs at $0.01 per million tokens.
Put a real client on it. Say 600 answered minutes a month on a plumbing line. The voice layer is about $30. Add a dozen decision questions firing across a six minute call, and you are still arguing about pennies. Then add telephony, the number, your own monitoring, and the hours you spend tuning the thing, and you land where every honest piece of voice math lands: the minutes are not the problem.
We charge $0.35 a minute on voice builds, which is a real markup over $0.05, and the reason is the stack around the minutes rather than the minutes. Our pricing page lists the wholesale tiers next to the monthly so you can do your own arithmetic on it.
The rest of what moved this month
Five things, in rough order of how much they should change your week.
Custom voices from a text description, free to create. On October 16, OpenAI released prompt-to-voice to all API customers. You describe a voice in words and get a voice. Creating it costs nothing, and speech generation bills at the normal GPT-Live rate. Nine new natural voices shipped alongside it. Separately, OpenAI's community announcement on October 15 described GPT-Live-1.1 rolling out with 13 new voices, better French, German, Hindi, Mexican Spanish, Korean and Japanese, and an explicit goal of reducing "checking..." pauses. That post says the API version was coming soon, so check availability before you promise a client a specific voice. If your clients have been asking why the agent does not sound like their region, this is the first month you have a real answer. We went through what moves the needle there in does an AI receptionist sound human.
Twelve hour prompt caching. Added October 20. Cache writes at that tier cost 2.5x the input token price. For an agent carrying a long, stable system prompt full of a client's services and price bands, this is a straightforward bill reduction.
Sandbox compute billed only on active turns. From October 9, OpenAI-hosted sandboxes can be pre-warmed, and compute bills during active turns rather than idle time between them. If you run anything agentic for clients, idle sandbox time was a quiet tax.
Anthropic cut cache reads and shipped a new small model. Claude Sonnet 5.5's cache read price dropped from $0.20 to $0.10 per million tokens on October 7, the same day Claude Haiku 5.5 launched. Same direction as everyone else: the cheap tier keeps getting cheaper.
And a deadline you should have in your calendar. Claude Sonnet 4.5 retires on November 30, 2026. If any client system you built is pinned to it, you have five weeks. Migrating off Haiku 4.5 also requires real changes, including dropping budget_tokens, temperature, top_p, top_k and prefill. This is the unglamorous half of running an agency, and we wrote about who owns it in what maintenance looks like after launch.
Where this does not help you
Four places, and I would rather you hear them from me than from a client.
It does not fix the lookup. Deciding to check the calendar faster is not the same as having a calendar connected. If your client's booking system has no usable API, or their tech availability lives in a group text, the agent now arrives at the dead end sooner. Most of the pauses that make AI phone systems feel robotic are integration problems wearing a latency costume.
It gives you a new way to be confidently wrong. A predicate returns a probability, not a fact. If you set your emergency threshold at 0.6 on partial transcript, you will sometimes dispatch on a sentence that was going somewhere else. Acting earlier means acting on less information, by definition. Pick thresholds per client and write down what happens on a false positive before you go live.
It adds plumbing you now have to get right. Out of order results. source_event_id matching. Waiting for session.updated. Permission scopes. None of this is hard, and all of it is one more thing that can break at 9pm on a Friday with nobody else to call.
You cannot evaluate it by feel. OpenAI's own voice agents guide tells you to report median and 95th percentile latency across similar calls, track backend time separately, and measure acknowledgements like "I'm checking" separately from the answer the caller actually needs. It also says to change one factor at a time and confirm a faster response did not hurt task success, tool correctness, or interruption handling. That is the right standard, and it is more work than most agencies do. If you are not measuring p95, you do not know whether your agent is fast. You know that it felt fast the three times you called it.
What I would change in your pitch this week
Stop demoing the agent answering. Everyone demos that, and two thirds of your prospects have now heard a bot answer a phone.
Demo the agent deciding. Call your own demo line, start a sentence that implies urgency, and stop talking halfway through. If the system is wired for this, the response that comes back is specific instead of a holding phrase. That is a ten second moment and it is worth more than a slide about model names.
Then ask one question on your next discovery call: when a customer calls about an emergency, what does whoever answers have to look up before they can give an answer? That is now a build spec. Before this month it was a reason to say the agent would take a message.
One thing not to do. Do not move a live client onto a new voice or a new decision path this week because it shipped this month. Run it on your own demo line first, make a dozen calls, look at the p95, then move one client who will tell you the truth.
FAQ
Does this mean an AI receptionist no longer pauses at all?
No. It means the pause can be filled with work instead of silence, and the work can start before the caller finishes. You will still hear a beat while the agent speaks its first words. The difference is whether the answer is ready when that beat ends.
Do I need to build this myself to sell it?
No, and I would not start there. Building it means your own GPT-Live session handling, your own delegation config, your own transcript trigger logic, and your own latency measurement. That is a real week of work before a single client sees it. Lumina builds run on our infrastructure with no API key for you to obtain, which is the whole reason the platform is $49 a month instead of a build invoice.
Is the Decisions API only useful for voice?
No. It is a general classifier endpoint, and it is good at the dull work that eats an agency's margin. Routing inbound form fills. Scoring lead quality. Flagging which of yesterday's 40 chat transcripts need a human to read them. The voice integration is the headline, but the cheapest win is probably a batch job you run nightly.
My client's trade is weird. Does any of this apply?
The mechanics apply to anything appointment based. What changes per trade is the question set: what counts as an emergency for a locksmith is nothing like what counts for a med spa. That is exactly what a vertical pack is for, and there are 100+ of them with the services, price bands and objections already written.
Should I tell clients which model their system runs on?
Tell them what it does and what it costs them. Model names change, retire, and get repriced, and a client who learned a model name in March will ask you about a headline in June. Sonnet 4.5 retiring on November 30 is a good example of why you want that conversation to be about capability rather than part numbers.
Your next step
Call the demo on our live demo section, pick a trade, and start a sentence about an emergency without finishing it. Listen to what the agent does with the half of the sentence it has. Then do it again on whatever you currently show prospects and compare the two.
If the comparison goes the way I think it will, start an agency workspace and get a demo under your own subdomain before your Monday calls. Builds get priced on a 30-minute call, so the thing you demo this week is the thing you can sell next week.
