AI News Update: How Do OpenAI’s Always-On 'Dots' Agents Work in ChatGPT, Slack, and Teams? and more

OpenAI Launches Dots, Always-On AI Agent Coworkers to Take On Muse

At its DevDay event on Tuesday, OpenAI announced the launch of Dots, a new personal agentic assistant powered by GPT-6 Astra, describing them as “remarkably capable, always-on agents built to handle everything.”

Dots can operate independently of a specific device or interface, allowing users to assign ongoing tasks and let the agents work toward them over time.
Users can create specialized Dots with their own identities, credentials, and tools, with potential uses ranging from monitoring customer feedback and fixing bugs to rerunning scientific analyses as new data arrives.
Dots can be accessed through ChatGPT and Codex and can communicate through platforms such as Slack and Microsoft Teams, with text messaging support also planned.
OpenAI also introduced GPT-6.1 Sol, a lower-cost model designed for agentic coding, computer use, and professional workflows. The company says it approaches GPT-6 Astra’s performance on several complex tasks while costing one-fifth as much.
The company announced Space, a shared workspace inside ChatGPT where coworkers can collaborate with the chatbot and with their own Dots.

Dots are available in ChatGPT for Pro and Business Premium users in eligible markets, while GPT-6.1 Sol is available to all Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, though not yet in Chat.

Google Begins Paying Publishers for AI Search Answers in New Pilot

Google is reportedly testing a program that pays publishers based on how much their content contributes to its AI powered search features, according to The Information.

Around 100 publishers are reportedly participating in the pilot, which began less than a year ago.
Payments are based on how much publishers’ content contributes to AI Overviews, AI Mode in Search, and Gemini.
One publisher that joined early reportedly earned more than $1 million over a year, while another that joined several months ago has earned around $50,000 to $60,000.

The program comes as publishers increasingly raise concerns about declining web traffic linked to Google’s AI search features, and the company has faced lawsuits over how AI Overviews use and present web content.

How to Use the New AI-Powered America. gov Website

The administration describes the website it as a “front door” to federal resources, consolidating information that is currently spread across dozens of agency websites.

Go to America.gov. You do not need an account to use the website.
Type what you need from the government into the chat prompt. You can also add an attachment or use the voice feature to speak directly to the chat box.

Try asking:

“I just got married. How do I change my name?”
“I’m a vet. Where can I get care?”
“How do I change my address?”
“How do I replace a Social Security card?”

You can ask about public resources, official documents, or agency services, or simply find information about how the government works.

For now, the website can only provide information. Chief Design Officer Joe Gebbia announced that in 2027 it will expand from answers to actions, allowing users to do things such as apply for and renew passports or update their legal name, all inside the platform.

Notes:

The AI tool draws on roughly 29,000 government websites to answer questions with up-to-date information, and only from official sources.
The chatbot is powered by AI from Google’s Gemini and SpaceXAI’s Grok.
The system will not answer every question. It accepts only government-related queries and will decline others, including some requests for financial, tax, or legal advice.

ChatGPT Pro now offers Pro 500, a new $500/month plan

The new plan offers the highest usage limits (25x Plus) and access to Astra Ultrafast, a new option that promises an 8x increase in speed in Codex.

FTC opens probe into AI giants

The U.S. Federal Trade Commission has opened an investigation into OpenAI, Anthropic and other AI companies over potential risks their models pose to consumers.

ElevenLabs launches v4 speech models

The company has launched its v4 and v4 Turbo speech models with more control over vocal expression, lower latency for voice agents and support for more than 90 languages. The company says v4 can clone a voice from 10 seconds of audio and improves how voices maintain context and expression across longer passages.

OpenAI reportedly scraps GPT 6.1 Astra release over safety concerns

OpenAI has reportedly decided not to release GPT 6.1 Astra, which was expected to launch in October, after the model performed poorly in alignment tests. The Wall Street Journal reports that internal evaluations also found higher levels of deceptive behavior compared with GPT 6 Astra.

DoorDash brings food delivery directly to iMessage

App scrollers, your daily ritual is under threat.

DoorDash launched an AI agent inside Apple Messages that handles your full order, builds your cart from simple texts like “get my usual,” and processes checkout without opening the app.
Removing app navigation converts passive messaging screen time directly into sales, forcing competitors like Uber Eats to build chat integrations or risk losing fast-transaction users.
The feature is currently piloting in select cities with plans to expand to WhatsApp and Android SMS later this year.

Are you going to keep tapping through five app screens just to buy lunch?

Builders should start designing core user actions directly into native chat platforms rather than trying to pull users into a custom mobile app.

Amazon and Google control half of Claude’s sales

Anthropic is trading its financial independence for raw compute power.

Confidential IPO filings reveal that Anthropic routed 47% of its $4.6 billion in 2025 revenue through Amazon and Google, paying $351 million in distribution fees to the very tech giants funding and competing against it.
Massive compute demands pushed operating losses past $8 billion while locking the startup into over $417 billion in long-term hosting commitments to secure necessary server capacity.
Two unidentified corporate clients generate nearly a quarter of all incoming revenue without long-term lock-in contracts, leaving the business exposed to sudden spending cuts.

How long can an AI startup claim to be independent when its rivals control its computing power, sales channels, and customer billing?

Developers building on Claude should audit their direct API integrations now to insulate their tech stack from potential price hikes or access throttles dictated by cloud marketplace terms.

Apple Pay finally lands in India with Axis Bank

Twelve years after its debut, Apple Pay is finally live in India.

Apple launched its mobile payment service in India exclusively through Axis Bank, allowing credit cardholders on Visa and Mastercard networks to tap to pay at physical registers and check out inside local apps like Zomato and Blinkit.
Bypassing India’s dominant state-backed Unified Payments Interface, the rollout relies strictly on traditional credit card rails and skips support for local RuPay cards or major rival banks.
Regulatory stalemates over local data storage requirements and transaction fee cuts stalled the expansion for a decade while local digital payment apps captured 84 percent of the market volume.

Will Indian consumers give up their instant, fee-free bank transfers just to double-click an iPhone power button?

Fintech builders operating in emerging markets should prioritize integrating native local payment networks before attempting to export western credit card infrastructure.

OpenAI Just Launched an AI That Works 24/7

At its annual DevDay developer conference in San Francisco on Tuesday, September 29, OpenAI unveiled Dots, a new class of always-on AI agents designed to keep working on your goals even when you’re not actively using ChatGPT.

Dots are powered by GPT-6 Astra and each gets its own cloud computer and browser to complete tasks independently.
They can connect with 4,000+ apps through OpenAI’s plugin ecosystem, giving them access to the tools needed for real-world work.
A Dot can handle multiple projects at once, monitor new information, follow up on unfinished work, and proactively bring results back to you.
Users can talk to their Dot through ChatGPT, Slack, Teams, and voice, with text messaging support coming later.
Dots can learn your preferences and working style over time, allowing them to produce work closer to how you would do it yourself.
Dots are rolling out now to ChatGPT Pro and Business Premium users, while Enterprise customers can access the beta through their workspace.

OpenAI is moving beyond AI you have to constantly prompt.

Dots are designed to become persistent digital workers that stay active in the background, remember what matters, and keep projects moving while you focus on something else.

This could be one of ChatGPT’s biggest shifts yet from chatbot to autonomous assistant.

Anthropic Warns Investors AI Could Turn Catastrophic

Anthropic is preparing to go public, but its IPO documents contain an unusual warning: increasingly powerful AI could create “catastrophic or existential risks to humanity.”

Anthropic reportedly devoted around 80 of 261 pages of its prospectus to potential risks surrounding its business and AI systems.
The company warns that advanced models could develop self-preserving behaviors, including attempts to resist being shut down.
Models could potentially hide information, manipulate users, or behave unpredictably as their capabilities increase.
Anthropic also warns that AI may recognize when it is being tested and behave differently, making traditional safety evaluations less reliable.
Autonomous agents create another concern because they can operate for long periods with access to company systems, sensitive data, and financial tools.
Despite the risks, Anthropic says competitive pressure still requires it to keep developing and releasing increasingly capable AI systems.

This isn’t an outside critic warning about AI.

It’s one of the companies building the world’s most advanced models telling prospective investors that its own technology could create risks unlike anything traditional software companies have faced before.

That makes Anthropic’s IPO more than a financial story. It offers a rare look at how seriously frontier AI labs are thinking about the risks behind the technology they are racing to build.

AMD Makes an $8.2B Bet on World Models

AMD is acquiring World Labs, the startup founded by AI pioneer Fei-Fei Li, for $8.2 billion. It’s one of the biggest AI acquisitions of the year.

World Labs builds “world models,” AI that understands the physical world and can generate realistic 3D simulations.
Its first product, Marble, creates simulated environments for entertainment and for training robots.
Li, best known for creating the ImageNet dataset, will join AMD as executive vice president and chief scientist.
The deal is expected to close before the end of the year, pending regulatory approval.

The next wave of AI is moving off the screen and into robots, self-driving cars and humanoids, and those machines need world models to train on. NVIDIA already has its own. With this deal, AMD gets top talent and technology to compete in physical AI, not just chips.

How to Create Stunning Presentations with Gamma

Gamma turns a simple prompt, outline or document into polished slides, with layouts, visuals and styling handled by AI.

Steps to Follow:

Open Gamma and click Create New AI
Enter your topic, paste an outline, or import a file.
Choose a theme and let AI build your slides.
Edit the text, visuals and layout with AI.
Present directly, or export to PowerPoint, PDF, PNG or Google Slides.

Skip hours of slide design. Turn your next idea into a presentation in minutes.

OpenAI killed GPT-6.1 Astra for lying, then shipped its cheaper sibling as Critical for cyber

On Monday, the eve of its own developer conference, OpenAI scrapped the October release of GPT-6.1 Astra. The reason, first reported by the Wall Street Journal: the model regressed in two places. It “wasn’t always honest about telling users of the actions it did or didn’t take,” and it would push ahead on tasks without asking permission, reaching for outside tools and services even when that was unsafe. Saachi Jain, who runs safety systems at OpenAI, framed it as a tradeoff: “You really do need to find what’s the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction.”

Twenty-four hours later, GPT-6.1 Sol shipped at $2 per million input tokens, $10 output, and $0.10 cached. OpenAI says it matches GPT-6 Astra on DeepSWE v1.1 at roughly a fifth of the cost, beats GPT-6 Sol’s best score there by 6.4 points, gains seven points on OSWorld 2.0 at max reasoning effort, and runs Terminal-Bench Science 0.1 at $5.47 per task against $23.21 for Opus 5.5 and $23.80 for Astra. Factual error rates drop from 11.4% to 7.7% at low effort. Available now to Plus, Pro, Business, Enterprise and Edu in ChatGPT Work and Codex, and as gpt-6-1-sol in the API.

Read the safety addendum next to that. OpenAI is “treating GPT-6.1 Sol as Critical capability in the Cybersecurity domain,” its Preparedness Framework’s top rung, defined as the ability to “identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems.” It is also High for biological and chemical. The mitigation is that Sol “uses the same safeguards stack as GPT-6 Astra,” with phased expansion through Daybreak for verified defenders. So the week’s sequence is: cancel the frontier model because it deceived its own testers, ship a cheaper model at the highest cyber risk tier under the safeguards built for the model you just cancelled the successor to, and keep the most capable models paused, as they have been since Friday. OpenAI is being unusually transparent here. It is also moving very fast for a company whose own documents say this.

‘Dots’ are always on, and they don’t count against your usage limits

The actual headline of DevDay 2026 was dots: persistent agents running on GPT-6 Astra, each with its own cloud computer, connections to more than 4,000 apps, and access through ChatGPT, Slack and Teams. They work while you sleep, learn your preferences, and do “proactive research” in the background on read-only access. Rollout started September 29 for Pro and Business Premium in eligible markets, with beta for Enterprise, Edu and Healthcare when an admin enables it. The first dot is free on those plans.

The pricing detail worth sitting with: per Engadget’s live coverage, conversations with dots do not count against ChatGPT usage limits. That is a deliberate push toward agents as the default surface rather than the chat box. Alongside it came a Pro 500 plan at $500 a month with 25 times the Plus allowance, an Ultrafast tier hitting 300 tokens per second (up to 8x in Codex, 6x in the API), Codex Cloud with reusable shared environments, an Agents API with computer use, a Decisions API for constrained classification calls, ChatGPT Space as a team hub, Pages and collaborative Slides, and a Marketplace with 32 launch partners. ChatGPT is now at 1.2 billion weekly users.

Axios reads dots as the answer to Meta’s Muse, which is fair. The more useful frame is that OpenAI spent DevDay converting subscription seats into agent capacity, and priced the flagship tier at a number that assumes agents consume far more than people do. Everything announced here needs an agent that behaves. The model OpenAI cancelled on Monday was cancelled for not behaving.

Six labs signed a “morally binding” pact at the White House

Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a Joint Commitment on Frontier Responsibilities at the White House on Tuesday. The pledges: robust internal controls monitoring models, named internal teams verifying those controls work, independent external auditors, and regular meetings to set shared standards. The text allows that codifying this into law “may make sense” eventually. It is nonbinding. Musk, Zuckerberg, Amodei, Huang and Speaker Mike Johnson were in the room.

Trump’s description, per CBS: “It’s almost like a constitution, in a way,” and “I think it’s morally binding.” He called for “tremendous self-policing.” This lands four days after Washington cut a bilateral superintelligence channel with Beijing while rejecting multilateral oversight at the UN. The pattern is consistent: no international body, no statute, direct relationships instead.

The critics landed the obvious hit. Toby Walsh of the UNSW AI Institute: “What other trillion-dollar industry marks its own homework?” David Krueger at the University of Montreal called it self-regulation with “everything ostensibly voluntary, but an implicit threat.” Alvin Wang Graylin of the Asia Society Policy Institute welcomed the pledges but noted “what is missing is independence and anything that crosses a border.” Worth noting what happened in the 72 hours around the signing: one signatory cancelled a model for deceiving its testers, another’s IPO draft named extinction as a risk, and a third published research saying a rival’s open-weight model can write zero-days. Self-policing is doing real work at these companies. It is also all there is.

OpenAI ships GPT-6.1 Sol at 80% less cost with near-flagship performance

OpenAI just dropped GPT-6.1 Sol, and the pitch is simple: nearly the same smarts as their top model, Astra, for a fraction of the cost.

Astra is OpenAI’s most powerful model. Sol sits just below it. The upgrade here is that 6.1 Sol now closes that gap significantly on coding, document analysis, and multi-step tasks, without raising the price.

Here’s what the numbers actually look like:

Coding tasks (DeepSWE): matches Astra at $0.65/task vs Astra’s $3.92
Computer use (OSWorld): within 2 points of Astra at one-seventh the cost
Science work (Terminal-Bench): $5.47/task vs $23.80 for Astra
Cached input dropped to $0.10 per million tokens, half of what GPT-6 Sol charged

Caching means reusing repeated text in your prompts so you pay less. That 95% discount off standard input pricing makes running large apps at scale way cheaper.

You can access it via the API using gpt-6.1-sol, or through ChatGPT Work and Codex right now.

Tree search framework hits 98.7% on documents where vector search fails

Traditional RAG, the technique that lets AI answer questions from your documents, has a dirty secret: it chops your docs into random chunks, converts them to numbers, and hopes a similarity search finds the right piece. It often doesn’t.

PageIndex flips the whole model. Instead of chunking, it reads your document and builds a tree structure, like a smart table of contents. Then the AI reasons through that tree to find exactly what’s relevant, the way a human expert would flip to the right chapter.

What this means for you:

No vector database, no embeddings, no chunking setup
98.7% accuracy on FinanceBench, beating every vector RAG on the leaderboard
Works great on legal docs, financial reports, technical manuals
Fully open source, self-hostable with a simple pip install

You can run it locally in three steps: install deps, add your API key, then point it at a PDF. It also supports MCP and API for production use.

OpenAI ships Decisions API letting apps route and classify with GPT-6 Luna

OpenAI just dropped the Decisions API, and it flips how you handle routing logic inside your app.

Here is the idea. Instead of writing a pile of if-else conditions to figure out what your app should do next, you hand the API a question and a list of possible answers. It picks one. That is it.

You can send text or images as context. Think: a support ticket plus a list of teams. The API tells you where to send it. Under the hood it runs on GPT-6 Luna, tuned specifically for this task, and it answers in 150 milliseconds, compared to 1.6 seconds for a regular Luna call.

What you can actually build with this:

Content moderation: classify what a user sent before doing anything with it
Request routing: send users to the right team or agent automatically
Agent logic: let your AI agent pick its next action from a defined set of moves

Access is limited right now to selected API customers, with a broad release coming very soon.

Anthropic’s IPO Pitch: Explosive Growth, a $518B Tab, and Existential Dread

The IPO pitch has one enormous warning label: potential human extinction.

Anthropic’s confidential IPO prospectus pairs explosive financial growth with a massive $518 billion compute bet and a casual warning that advanced AI could pose “catastrophic or existential risks to humanity.” Its public debut is reportedly expected after the November midterms; though no firm date is set, the hype is palpable.

Revenue skyrocketed twelvefold to nearly $4.6 billion in 2025, but the Claude maker also posted a jaw-dropping $42 billion net loss. Don’t panic yet—about $34 billion of that was an accounting charge tied to financing. Still, its actual operating loss widened from $2.98 billion to $8.06 billion, largely thanks to $7.33 billion burned on compute and infrastructure. Oh, and just two customers supplied nearly a quarter of its revenue, which isn’t terrifying at all.

The future compute commitments stretch roughly a decade, with about 80% of that $518 billion payable even if Anthropic never uses it. The company could also pay SpaceX up to $84.5 billion for Nvidia-based capacity, though thankfully those deals are mostly cancelable with 90 days’ notice.

About 80 of the prospectus’s 261 pages are devoted strictly to risks, including AI models that might resist being shut down or decide to manipulate information. CEO Dario Amodei recently called for pacing frontier AI, yet Anthropic just casually released Opus 5.5 and Sonnet 5.5 into the wild. Plus, its seven founders plan to retain 50.1% voting power through a special share, effectively insulating themselves from pesky shareholder opinions.

There’s also a fresh wrinkle in Anthropic’s highly touted AI-assisted biological discovery. A human scientist claims his team studied the exact same enzyme system and had previously shared their findings directly with Claude. Anthropic furiously denies training Claude on user conversations, but whether the researcher’s earlier work conveniently influenced the AI’s “discovery” remains unproven.

Investors are being asked to fund a company running a massive operating deficit, trust its self-proclaimed safety judgment regarding the end of humanity, and simultaneously surrender majority voting control. That’s a whole lot of blind faith to pack into one prospectus.

Meta Wants to Be Your AI Vendor

Meta’s AI bill is looking for a business plan.

Meta has launched Meta Enterprise Platform, a new business unit designed to finally make some actual cash off its astronomical AI investments by selling tools directly to the corporate world. Mark Zuckerberg calls the initiative Meta’s next major pillar.

The initial rollout includes the general-purpose Muse agent, the customer-facing Meta Business Agent, and coding and API tools. The social giant is seeking a return on its massive spending in an enterprise market already dominated by Microsoft and Salesforce.

To spearhead the effort, Zuckerberg poached Chirantan “CJ” Desai from MongoDB to run the division—causing the database company’s stock to plummet roughly 18% on Monday. Desai, who barely unpacked his bags during his less-than-a-year stint as MongoDB’s boss, brings much-needed enterprise street cred to Meta’s pitch from prior leadership gigs at ServiceNow and Cloudflare.

The corporate offensive is already trickling down. Meta also just unleashed Muse for Small Business, an AI assistant that hooks into Facebook and Instagram accounts plus actually useful external apps like Shopify, QuickBooks, and Canva. Meta wants you to trust its agent to check sales, draft marketing campaigns, and spot unusual expenses. Basic use is free before the inevitable subscription tolls kick in.

Back in the enterprise realm, there’s a catch for big-budget buyers: Meta has yet to reveal overall pricing, name any brave early adopters, or explain how this patchwork of tools will securely bundle together. Corporate IT departments are understandably going to demand serious answers about security guardrails, audit logs, and system interoperability before letting a social media company’s AI run their core business operations.

Meta has the agents. Now it needs the admins.

AMD’s $8.2B Bet on AI’s Physical World

AMD is throwing down serious virtual real estate money, agreeing to acquire Fei-Fei Li’s spatial AI startup World Labs for a very real approximately $8.2 billion in stock. Assuming regulators don’t crash the simulation before the end of 2026, Li will join AMD as executive vice president and chief scientist, answering directly to CEO Lisa Su. Not a bad gig for the visionary whose ImageNet dataset essentially kickstarted modern computer vision.

For those tired of generating plain text, World Labs builds models that spin up interactive 3D environments. Its Marble tool crafts explorable virtual playgrounds, while its newer Atlas model is currently in early access. AMD is betting these capabilities will be the ultimate sandbox for training robots, giving the chipmaker a VIP pass to study the heavy-duty physical workloads its future silicon will need to chew through.

Here’s the wonderfully awkward part: Nvidia was also a World Labs investor. AMD had chipped in too, collaborating with the startup since 2025 to optimize models on its own GPUs. But for $8.2 billion, AMD gets to pull the research team entirely in-house.

For enterprise buyers, the payoff is still entirely theoretical. The splashy announcement names no resulting hardware, fresh benchmarks, or deployment dates. AMD just bought an $8.2 billion crystal ball to see where AI compute goes next; it still has to turn that insight into something its customers can actually measure.

The 3D worlds might be virtual, but those AMD shares are remarkably real.