The Single Letter on the Pricing Page: OpenAI's 'Always-On' Assistant Claim

The single letter on the pricing page

One letter, one large claim: the always-on assistant line that appeared on OpenAI's pricing screen may herald a helper that keeps working after you close the chat window. The screenshot shared by Jake Boggs on September 25 was viewed hundreds of thousands of times, and combined with testingcatalog's findings it became more than an ordinary rumor. Still, there is no product — only a line on a payments page and a configuration record embedded in the client.

The details of the screen stand out because the line sits among paid features. On the $100 plan labeled Standard, it stands next to the most capable models, extra Codex usage, maximum memory, 100 GB of storage, and early access. As the speaker reports, the list covers everything in Plus and adds more. So 'o' is staged not as a separate product page but as a subscription benefit. That placement suggests the company takes it seriously, yet it proves nothing about the final price.

The second piece of evidence is more technical. The dynamic configuration fragment shared by testingcatalog shows 'o' as the display name and '-o' as the email identity suffix. The record is evaluated per account, and the feature reads as switched off for the account that captured it; the infrastructure sits in the client, but the switch is unflipped. It is too early to infer a mailbox from this: the suffix could be a real address or a simple alias. Even so, customer-facing wording and an engineering record giving the same name weakens the coincidence theory.

The September 29 window

Beyond that lies a map of gaps. Which tasks the assistant would take on, which permissions and memory behavior it would run with, how it would be scheduled, and who would get access are all unknown. The $100 tag says little on its own; separate reports mention a paused $200 plan and a new $500 top tier, but none of it is confirmed on an official page. The speaker draws an honest line here: a pricing screen does not explain what a product does.

The date to watch is fixed: DevDay at Fort Mason in San Francisco on September 29, with Sam Altman giving the opening keynote; the schedule is confirmed on openai's event page. openai mentions no 'o' anywhere on that page, so the announcement expectation rests on testingcatalog's probability assessment. Consumer features being staged alongside developer tools lately makes the expectation reasonable — but not a guarantee.

Behind the curtain sits the AON puzzle. The name previously surfaced around OpenAI's custom-agent work for Workspace accounts; one reading says 'o' becomes the consumer face of the same infrastructure. Another reading casts AON as the answer to Grok Bot. Both narratives cannot hold at once, and the available records do not say which one stands. Whether these names are internal codenames, separate products, or versions of one system will only be settled by an official document.

Signal or noise?

Signal must be separated from noise, because the scene is crowded. Staff posts saying 'o no' and 'o yes' sit somewhere between joke and teaser; whether Plus subscribers get access went unanswered. No documentation ties the single-letter @o handle to the company, and no link connects the over-$300 ring-device rumor to the teaser clips. Claims of 63-language support, a Cerebras-powered fast mode, and access across all Pro tiers circulate with no primary source either.

Everyone in the same race

The race itself, however, is concrete. According to reuters, Meta launched Muse in the US on September 8: sending email, arranging travel, paying through Stripe Link, wrapped in an isolated virtual machine with confirmation steps. According to venturebeat, Grok Bot entered early beta in August; with its own computer it works inside applications, sustains tasks around the clock, and starts at $120 per seat. Both sell the same promise: a helper that does not stop when you close the window. That picture explains why OpenAI might hurry.

On the speed front the picture is more measurable. openai's Ultrafast preview runs GPT-5.6 Sol up to 14 times faster than standard processing, generating 750 tokens per second; access stays limited to selected customers for now. According to cnbc, the Cerebras deal points to 750 megawatts of compute through 2028 at a scale above $10 billion. A standard, fast, and ultrafast selector in the Responses API also signals expansion around DevDay. Speed is a tier sold for money; the assistant is still a promise.

Microsoft, for its part, has already made its announcement. According to microsoft's blog, the company rebuilt Copilot on September 25 around the Home, Code, and Autopilot trio: Home and Code reach the Frontier program in the coming weeks, Autopilot expands to private preview at month's end. In the supplier-review example you hand it a goal, it builds the schedule, sets the meeting, and owns the follow-ups — behaving like a participant you can tag in Teams and Outlook. Billing splits in two: daily use in the subscription, long-running agent work metered by usage. The FinOps discipline exists for exactly that.

Google closes with a different door. According to thurrott, the Gemini live avatar is served through Gemini Enterprise; a face generated alongside speech promises lip sync and 97-language support, and the hotel-desk demo keeps talking while calling tools in the background. Generated audio and video carry an invisible SynthID watermark, and custom faces stay behind an allowlist. The real test is identical on all three fronts: remembering the goal, noticing change, and knowing when to ask.