OpenAI's New Agent O Changes ChatGPT Forever transcript

OpenAI may have accidentally put its next major product on a subscription page before announcing it. The name is just O and the description is your always on assistant. That is a pretty big thing to slip into a pricing screen days before dev day because it suggests something that keeps working after you leave chat GPT. The screenshot came from Jake Bogs, who posted it on September 25 asking whether he had missed a launch or found a Devday leak.

One capture recorded over 417,000 views. The thread later showed 620,500. People immediately noticed that this wasn't buried in some technical document. It was sitting among benefits people would pay for. The screen showed a $100 Chat GPT Pro plan labeled standard, described as Frontier Intelligence. available in one tier. Under everything included in Plus, it listed the most capable models, the always on assistant, more work and codec usage, maximum memory, 100 GB of storage, and early access to features.

That doesn't establish final pricing or eligibility, but it places O inside an actual subscription pitch. Then testing catalog found another clue. Configuration containing the display name O and an email suffix of O. Now an assistant with a possible email identity is more interesting than another chat button. You could potentially assign work to something that remains reachable while you're elsewhere. But those fields alone don't prove it can read your inbox, send messages, or operate its own mailbox.

The suffix could identify an alias or something else entirely. The configuration screenshot also contains accountbased settings and fields marked false. Those details suggest a feature being controlled behind the scenes, potentially inactive for the account shown. But the screenshot alone doesn't establish exactly how the feature works. Other nearby settings have no clear connection to the assistant, so they don't tell us much about what it can actually do.

Still, you have customerf facing wording and configuration pointing toward the same name. That's more substantial than somebody posting a mysterious emoji. What remains missing is the product itself, supported tasks, permissions, scheduling, memory behavior, availability, and whether that $100 listing reflects what OpenAI will actually sell. September 29 is the date everyone is watching. Dev Day is scheduled for Fort Mason in San Francisco with Sam Alman's keynote at 10:00 a.m.

Pacific, 100 p.m. Eastern, and 6:00 pm London time. The event is scheduled and O announcement isn't confirmed. Testing catalog identifies it as a possible launch window, which is an important distinction when headlines already make the reveal sound guaranteed. There's also AON, a name previously associated with OpenAI's always on agent work and custom agents for chat GPT workspace accounts. One interpretation is that O becomes the consumer version of related infrastructure.

Other reports describe the relationship differently, including AON as an answer to Grockbot. We don't yet know whether these are separate projects, internal names, or different versions of the same system. And yes, people are reading into OpenAI posts. Tibo Thibo Sdio used oh no and oh yes around a service disruption. His post said paid users would have usage limits reset across codecs and chat GPT work and mentioned a spare codeex helping when things go down.

That could be playful teasing or just ordinary wording. Commenters debated the name and whether plus subscribers would get access. Nobody answered the subscription question. The suspended account with the handle O onx has also attracted attention, but there is no documented ownership linking it to Open AI. Then there's the rumored donut-shaped device reportedly above $300 and Amberino's teaser showing a dark rectangle above Earth.

People are connecting the shape to the name. Open AI hasn't established that connection, and a teaser isn't enough to turn this into a hardware launch. Some coverage goes further, claiming a Tuesday release, access across every pro plan, 63 languages, and a Cerebrra's powered fast mode. Those details remain unconfirmed. The same applies to the rumored $500 plan. Testing catalog separately mentions a leaked Pro Max subscription, but that doesn't settle its price or how this assistant would fit into it.

The appeal is pretty straightforward, though. Give an assistant recurring research, monitoring or communications work and have it continue without reopening the conversation every time. Those are potential uses, not demonstrated O features. The real test will be whether it remembers the objective, recognizes when circumstances change, and knows when it needs your input. The competitive backdrop explains the urgency. Reports describe Meta launching Muse on September 11 for email, payments, and health related tasks with a Muse Charm companion planned for December.

Other developments in this space include XAI's September 5 Grobbot announcement, Anthropics Conway and leaked scheduling features, Claude Tag in Slack, and Google's roundthe-clock Gemini Spark. These products don't necessarily offer the same capabilities, but the direction is similar. agents that stay involved beyond a single conversation. Open AAI also has another possible dev day development that is easier to measure speed.

Testing catalog spotted a hidden selector being prepared for the responses API playground offering standard fast and ultraast processing. A wider roll out around the event looks plausible although the selector appearing doesn't establish when everyone gets access. According to testing catalog, OpenAI already previewed ultraast with GPT 5.6 Soul at up to 750 output tokens per second or up to 14 times faster inference than standard.

Access is limited to selected customers and Cerebras powers the mode. Those are maximum reported figures, not a promise that every request or complete workflow finishes 14 times faster. GPT6, Saul, Luna, and Astra have since entered the picture with reports also describing API price reductions of half, but support across every GPT6 model hasn't been confirmed. Developers may eventually choose processing speed according to the job, paying more where latency affects productivity or revenue.

Actual ultraast economics still need to be established. Behind that sits a reported 750 megawatt Cerebras's partnership with capacity arriving in stages through 2028. Testing catalog has also spotted platform onboarding tiers including accelerate for production workloads. Together the findings suggest a developer event with substantial attention on infrastructure compute options and running applications at scale. Microsoft meanwhile has published a much more detailed description of its own persistent agent.

Its September 25 announcement introduces a rebuilt co-pilot around home code and autopilot. Home and code begin arriving through Frontier in the coming weeks, while autopilot expands into private preview at the end of the month. Autopilot was previously called Scout. You give it a name, role, and goal. And Microsoft says it can watch channels, follow threads, run recurring work, and resume projects days later. Its supplier review example includes building a schedule and workback plan, preparing meetings, handling follow-ups, and contacting stakeholders for updates.

It runs in the cloud inside the organization's tenant with its own identity, memory, computer, and workspace. Microsoft IQ supplies business context while permissions, auditing, and governance define its boundaries. It can appear in Teams, Outlook, chats, channels, and documents where colleagues can mention it like another participant. The promise is continued work while you're asleep with updates that keep you informed.

Home brings chat and co-work together, showing recent activity, suggestions, and unfinished work. Chat handles quick questions, lookups, and drafts. Co-work takes delegated assignments through to completed results, including proposal responses, launch kits, customer briefings, and financial close packages. Eventually, Copilot is supposed to route requests automatically into chat, co-work, or code. Office is also moving directly into that experience.

Word documents, Excel workbooks, and PowerPoint presentations remain real, editable files, synchronized with the Office apps, and available for live team collaboration. You can refine them alongside the conversation or mention a teammate whose changes appear as they work. PowerPoint gains automatic brand consistency, slide and image adjustments, and custom templates. Excel explains what co-pilot or collaborators changed and why, and recommends charts.

Reusable skills brings specialized workflows into applications, including finance and Excel and legal work in Word. Code lets people describe trackers, dashboards, automations, workflows, desktop widgets, or internal cloud applications in everyday language. It uses the underlying technology behind GitHub Copilot, runs in a sandbox, can be hosted within the tenant, draws context from Microsoft IQ, and automatically synchronizes plugins.

Developers can continue using GitHub C-Pilot with closer platform connections. Code reaches Frontier at month's end with broader availability in the following weeks and previews for Microsoft 365 Premium and Pro later this year. Copilot managed runtime already in preview provides it governed hosting inside Microsoft 365 including shared apps connected to live data. It supports co-work and co-pilot studio creations will be accessible inside code and is opening to thirdparty and professional developers.

Microsoft IQ combines organizational knowledge and business operations. Fabric IQ adds data context including more than 20 million PowerBI semantic models and is generally available in chat and co-work with code integration coming through Frontier. Dynamics 365 and Power Platform grounding enters public preview over the next month allowing a proposal to incorporate deal history and support tickets directly. A unified plug-in registry is rolling out now with general availability across surfaces expected in the coming weeks.

It combines Microsoft Partner and custom plugins containing skills, connectors, and other capabilities. It approves them centrally. Developers publish once for supported co-pilot experiences. The billing split deserves attention. A user subscription license covers everyday co-pilot use across chat and office applications, model selection, and new capabilities. Auto routes requests by weighing accuracy, speed, and cost. co-work autopilot longer running agent work and frontier models such as Astra and Fable use usage based billing instead.

Microsoft calls its spending management approach phops for AI. Agent 365 cost controls expand beyond co-work and work IQ APIs to code and managed runtime with co-pilot studio agent support planned for October. Admins get spending policy APIs, credit request approval workflows, and model family restrictions that also constrain auto. Leaders can assess co-work outcomes while users see credit consumption, balances, and history directly in Copilot.

Coming next today turns home into a personalized view across email, calendars, teams, meetings, and tasks. It can prepare drafts, proposed schedule changes, and surface follow-ups with context. Private preview starts in October with Outlook and Teams versions planned. A separate co-pilot preview in Teams arrives by months end, letting people mention co-pilot directly in a conversation. It uses shared conversation context and permissions to explain decisions and prepare crossf functional reviews with dependencies, owners, questions, and unresolved decisions.

Microsoft points to Ignite November 17 through 20 in San Francisco and online for more. Google's announcement gives its agent a different kind of presence, a face that responds while you talk. Gemini 3.8 Live with live avatar follows the previous week's Gemini 3.8 live launch and is available through Gemini Enterprise. It pairs native live dialogue with nearrealtime video generation, producing speech, lip movements, expressions, and conversational turnt taking together.

The model processes visual and audio inputs simultaneously so it can respond to what it sees and hears. Google's demonstrations show different characters with distinct appearances, voices, and expressive styles. Suggested applications include customer service and interactive walkthroughs where the visible persona becomes part of the conversation. It also supports asynchronous tool execution. In the hotel demonstration, the avatar handles a guest check-in while calling tools behind the scenes rather than pausing the entire exchange while information loads.

Background operations and active dialogue can overlap, which should make those interactions feel less interrupted. Google says it supports 97 languages with native speech-to-pech synchronization and lip movements and expressions adapting as languages change. The company also says those switches preserve video fidelity without visual drift. That's a substantial claim for an animated character that needs to remain consistent throughout a conversation.

Businesses can choose preset avatars or create one from a highquality reference image preserving likeness, styling, or character identity. Custom creation currently requires enterprise allow listing. Generated audio and video contain imperceptible synth ID watermarks intended to help detect AI generated material and reduce misattribution. Google provides a model card covering its safeguards with access through Gemini Enterprise and implementation details in the live API documentation.

Thanks for watching and I'll catch you in the next one.