Between September 2 and October 7, 2026, OpenAI put out about thirty announcements. Reading them all is a full-time job, and it isn’t yours.
Here’s the short version. Most are price tags and packaging on things you could already buy. A mid-size model at $2 USD per million input tokens (OpenAI’s pricing page) lands where third-party trackers put Anthropic’s mid-size model. A faster image generator, new subscription tiers, coding tools. None of that moves a small business.
Three releases do. Each opens a door that was shut in August. They also share a trap that none of the launch posts mention, and it has nothing to do with price: the machines they run on aren’t in Canada.
Skip These, For Now
| Announcement | Reason to wait |
|---|---|
| Sora 2 and the video API | Shut down September 24, 2026 (reported). OpenAI sells no video model through its API. |
| Hosted agents (Agents API) | Beta, United States only, keeps its state until deleted, can’t run under zero retention (OpenAI’s data controls page). |
| Dots, the always-on agents | One ships inside the plan and gets configured in the product. No third party needed. |
| OpenAI’s partner marketplace | Open to Enterprise customers only. |
| Older image models | gpt-image-1 is retired on October 23, 2026 (reported). If software you rely on calls it, ask the vendor about its migration. |
1. The Answer That Comes With Odds
The Decisions API went into public beta on October 6, 2026. Instead of writing back, it hands you a probability, a pick from a list, or a score on an ordered scale, and one call can carry up to 200 questions about a single text or image (OpenAI’s guide). Ask whether an email is about billing and the reply might be “yes, 94%.”
That figure is what you’ve been missing. A regular language model gives you prose; you can’t draw a line through prose. With a number, you can: anything at 90% or above gets handled automatically, anything below goes to a person.
On price, third-party sites report input-only billing at $0.10 USD per million tokens, or about $0.00004 USD for a 400-token decision (aifreeapi.com). An independent tester paid $0.041 USD for 1,500 decisions (vincentvandeth.nl). Checking every item stops being a cost question.
Whether the odds can be trusted depends on the job. Two tests disagree:
- A 450-text classification test found it right 95% of the time. Every one of its 415 answers at 90% or higher was correct. Anthropic’s small Claude Haiku 5.5 slipped less often (15 errors in 600, against 30) but offered no probability (same source).
- A 3,600-problem logic test found that when the underlying model claimed 99% or more, it was right 68% of the time. Once an answer took five chained inferences, accuracy dropped to chance (anth.us). The author favours a rival product. Weigh it accordingly.
Our reading: trust it where the answer sits in plain view in the input, such as which department a message belongs to or whether an item shows up in a photo. Don’t use it to gate multi-step reasoning. We found no independent test on photographs, and the service is still in beta, listed by OpenAI under United States processing.
2. A Voice That Picks Up the Phone
GPT-Live-1 reached the API on September 10, 2026. It listens as it talks, lets you interrupt, handles phone-quality audio, and passes tool work to a second model (OpenAI’s announcement). “Full-duplex” is the jargon. The effect is a call that sounds like a call.
Third-party pages report $0.05 USD per minute, billed by the second, with your backend on top (Coursiv). That puts a three-minute call near $0.15 USD. OpenAI’s own page isn’t the source for that number, so check it before you plan around it. A flat rate beats the audio-token billing earlier voice models used, because you can predict the monthly bill.
Three things to watch:
- Quebec French is untested. We haven’t heard this model speak it, with a Sherbrooke accent or any other. Don’t assume it’s good or bad. Have three local people try it before you commit.
- A voice is personal information, and this one is processed in the United States. Stored recordings last 30 days when storage is on, and zero retention switches storage off (reported by Wiro).
- Callers deserve to know what they’re talking to. Say that it’s an automated assistant.
3. ChatGPT as a Storefront
ChatGPT now offers two ways for a business to be seen.
Ads. They show to Free and Go plan users. OpenAI announced an Ads Manager plugin on September 16, 2026, which builds campaigns from plain-language instructions, along with Shopify and HubSpot apps (announcement). On October 5 came an image ad that appears next to image generation, labelled as an ad and separate from the picture; OpenAI calls it a test (announcement). Third-party guides say self-serve buying opened in May 2026 with no minimum spend. OpenAI’s post on its ad tests includes Canada (OpenAI), and guides report Canadian businesses are eligible.
Plugins. Since September 29, 2026, any verified organization can submit a plugin with its own interface, shown to ChatGPT users (submission page). A shop can put its catalogue in one, or a business its bookings.
The Ads Manager is free to use. A third-party blog reports a recommended bid of $3 to $5 per click (Flyweel). No public price turned up for agencies that run these campaigns.
Where it gets sticky:
- Paying subscribers never see the ads, so you only reach the free tier.
- Reports list excluded categories: business software, professional services, and agencies (Witscode). Accountants and consultants should check first.
- A plugin’s checkout covers physical goods only and finishes on the merchant’s site (OpenAI).
- Visibility is OpenAI’s to give. Developers on OpenAI’s forum say in-chat search has missed some plugins since September 9, though the directory still shows them (forum thread). Unconfirmed, but a fair warning about building on someone else’s ground.
Also worth noting: the new tools (plain-language campaigns, Shopify apps) strip out steps, including the ones an agency would bill you for.
All Three, Side by Side
| What you get | Published price | Biggest catch | |
|---|---|---|---|
| Decisions API | A probability, choice, or score you can set a threshold on | $0.10 USD per million tokens, input only (reported) | Solid on simple judgments, weak on chained reasoning; no photo tests |
| GPT-Live-1 | Phone conversation that handles interruptions | $0.05 USD per minute (reported) | Quebec French untested; voice processed in the United States |
| ChatGPT ads and plugins | Visibility with Free and Go users | No minimum; $3 to $5 per click (reported) | Excluded categories; checkout for physical goods only |
Ask Where the Server Is
Before you compare prices, ask this. None of the three runs its models in Canada. OpenAI’s Canada region holds data and does no processing. Regional processing is limited to the United States and Europe (OpenAI, data controls), and the Decisions API sits under United States processing.
Here’s what that means under Quebec’s privacy law, in plain words. Since September 2023, an organization sending personal information outside Quebec must assess the privacy risk before it does (Commission d’accès à l’information). A name, a phone number, a customer’s voice, and a photo of a person all qualify. The border that counts is Quebec’s, not Canada’s.
So:
- If what you send identifies no one, a product catalogue for example, the question doesn’t come up for that transfer.
- If a customer types their name into a chat, speaks on a call, or has a photo analysed, do the assessment first. No assessment, no sending.
A rate per minute or per click is silent on that step. It belongs in the real cost.
What We Didn’t Check
We opened no OpenAI account and made no paid call for this piece. Prices for voice, the Decisions API, and some models come from third-party pages rather than OpenAI’s price page, and we flagged each one. The OpenAI announcements cited here come from OpenAI’s own pages. Still open: how accurate the Decisions API is on photographs, how GPT-Live-1 handles Quebec French, and what agencies charge to manage ads.
FAQ
What is OpenAI’s Decisions API, in plain terms?
A service that answers with a number instead of a paragraph. You ask a question about a text or an image, and it returns a probability, a choice, or a score. You set a threshold: below 90%, a person looks. It has been in public beta since October 6, 2026, running on the gpt-6-luna model.
What does OpenAI’s new phone voice model, GPT-Live-1, cost?
Third-party pricing pages report $0.05 USD per minute, billed per second, plus the cost of whatever system does the work behind the call. A three-minute call comes to about $0.15 USD. That figure comes from those pages, not from OpenAI. How well it speaks Quebec French is untested.
Can my business advertise inside ChatGPT?
According to third-party guides, yes: Canadian businesses are reported eligible, with no minimum spend. But the ads reach only Free and Go plan users, some categories are reportedly excluded (business software and professional services among them), and OpenAI still calls the image format a test.
Does OpenAI process my data in Canada?
Not for these three announcements. OpenAI’s Canada region stores data but doesn’t run the models. Regional processing exists only in the United States and Europe. If customers’ personal information goes to OpenAI, it leaves Quebec, and Quebec’s privacy law calls for an assessment before it does.
If you’d like help deciding which of the three deserves a trial in your business, the consultation is free.