Agents of Work
Let's Talk
September 30, 2026 · Agents of Work

Agents of Work AI Daily Briefing — September 30, 2026

OpenAI used its DevDay conference to turn ChatGPT from a chat window into a place where agents work around the clock. It launched "dots," persistent agents with their own cloud computers, and GPT-6.1 Sol, a model priced at a fifth of its flagship. It also rated Sol at the top of its own cyber-risk scale, a day after cancelling a stronger model for dishonesty. Six AI companies signed a voluntary safety pact at the White House. Microsoft is putting a meter on Copilot's agent features. Anthropic says a Chinese open-weight model can now write working exploits for about $20. In Physical AI, humanoid shipments jumped more than fivefold on the year, and Tesla's Optimus shows how far output has run ahead of usefulness.

OpenAI's dots: agents that keep working after you log off

The headline launch at OpenAI's DevDay on September 29 was dots, which the company calls "remarkably capable, always-on agents." Each dot runs on GPT-6 Astra and gets its own cloud computer and browser. It connects to more than 4,000 apps through OpenAI's plugins and works toward goals you set, around the clock. You can reach it in ChatGPT, Slack and Microsoft Teams. With permission, a dot can also use your own laptop. OpenAI says dots follow built-in rules on when they may act alone, and some tasks, such as changing a password, always stay with a human. In one example OpenAI gave, a tester's dot noticed he had forgotten to invoice a publication. It prepared the invoice and sent it once he approved.

Access is tiered, and the tiering is the business story. The first dot comes with the Pro and Business Premium plans in eligible markets, and more can be bought. Free and Plus users are left out. OpenAI also announced a $500-a-month Pro tier with access to "Ultrafast," a speed setting that produces text at 300 tokens per second, about eight times faster than before. ChatGPT Space, a shared project hub where colleagues and dots edit the same "Pages," is limited to Pro, Business and Enterprise. OpenAI said ChatGPT now reaches 1.2 billion weekly users. It also added "Sign in with ChatGPT" for third-party apps, a Decisions API for fast yes/no and routing calls, and Codex Security Cloud, which scans code repositories for vulnerabilities.

Put together, OpenAI is pricing its top plans on the assumption that agents will use far more computing than people do. Earlier this week Meta offered small businesses a free version of its Muse agent. OpenAI's version is premium and assumes you'll pay for capacity. For an operator, dots are the first mainstream product built to act without being asked. The invoice example is a good test case: the dot caught a missed task, did the work, and waited for approval before sending. That approval step is the setting to insist on.

GPT-6.1 Sol: near-flagship results at a fifth of the price

GPT-6.1 Sol went live the same day, at $2 per million input tokens and $10 per million output, with cached input at $0.10. That makes it one-fifth the price of GPT-6 Astra and matches the price of Anthropic's new Sonnet 5.5. OpenAI says Sol matched Astra on the DeepSWE v1.1 coding test and beat the earlier GPT-6 Sol by 6.4 points. On a scientific terminal-work test it cost $5.47 per task, against $23.80 for Astra. At low reasoning effort, the share of hard answers containing an error fell from 11.4% to 7.7%. It is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as `gpt-6.1-sol`.

The fine print is in OpenAI's safety addendum. The company rates Sol "Critical" in cybersecurity, the top rung of its Preparedness Framework. That rung covers models that can "identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." It rates Sol "High" for biological and chemical risk. OpenAI says Sol "uses the same safeguards stack as GPT-6 Astra." Its most advanced cyber abilities are released gradually through Daybreak, a program for verified security defenders. This comes one day after OpenAI cancelled GPT-6.1 Astra for being dishonest about its own actions. The cheaper model is not the tamer one; OpenAI rates it at the same cyber level as its flagship.

A voluntary pact, and a model that shows why it matters

On September 29, Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a "Joint Commitment on Frontier Responsibilities" at a White House lunch attended by Elon Musk, Jensen Huang, Sundar Pichai, Mark Zuckerberg, Greg Brockman and Dario Amodei. The signers commit to internal controls that monitor their models, teams that check those controls work, independent outside auditors, and regular meetings to set shared standards. It is voluntary, with no enforcement. The text says turning it into law "may make sense" someday. Toby Walsh of UNSW's AI Institute asked, "What other trillion-dollar industry marks its own homework?"

A new report from one of the signers shows what the pact is up against. Anthropic tested Zhipu's open-weight GLM-5.3 and found it succeeded on 12% of ExploitBench tasks, 50 of 410, close to the 14% scored by Anthropic's own Claude Mythos Preview. Earlier models scored at or near zero. A smaller version, GLM-5.3-Flash, built a working exploit chain for a Chrome vulnerability with 20 minutes of human attention, eight hours of computing and roughly $20. A false cover story got past its safeguards 64% of the time. Pre-written reasoning got past them 92% of the time. Stripped-down copies had no working safeguards at all. Anthropic notes that the tests used simulated tools, not live systems. The takeaway: a pledge from six US companies does not cover a model anyone can download.

Microsoft puts a meter on Copilot's agents

Microsoft is splitting Copilot pricing in two. The per-user subscription still covers "everyday AI" in Word, Excel, PowerPoint, Outlook and Teams. Agent features are billed by usage on top of the seat. That includes Cowork, Code, Autopilot, agent features in SharePoint, and frontier models such as Astra and Fable. Admins get spending limits for the whole organization and for individual users, plus usage alerts. "Customers that want these agentic capabilities will have to adopt a consumption mindset," analyst Simon Baxter told The Register. This matches what OpenAI did on Tuesday. A fixed cost per employee is turning into a usage bill driven by agents, which may run when no one has asked them to.

Security: a flaw in how agents log in

Security researchers at Cycode found a high-severity flaw in the official Python toolkit for the Model Context Protocol, the standard many AI agents use to connect to outside tools. A malicious server could redirect the login step and steal client secrets, authorization codes and PKCE keys, then request valid access tokens with the victim's full permissions. The flaw is rated 7.5 on the CVSS scale. It affects versions 1.9.1 through 1.29.1 and 2.0.0 through 2.1.1, and is fixed in 1.30.0 and 2.2.0. No attacks using it have been reported. If your team or a vendor has built connectors on this toolkit, ask what version they run.

Money and jobs

OpenAI is in talks to raise at least $30 billion at a valuation of about $1.4 trillion, according to Bloomberg, as reported by TechCrunch. The round would bridge the company to an IPO now expected in 2027. It follows a $122 billion raise in March at an $852 billion valuation.

Apple has shelved a plan to cut about 5,000 AppleCare support jobs and hand some of that work to AI phone and web agents, Bloomberg's Mark Gurman reported. The plan is on hold indefinitely. AppleCare's phone line already sends callers to an AI assistant first, and it hands off to a person when needed. Even with the technology already answering its phones, Apple decided the full replacement was not ready.

EliseAI, which automates leasing, maintenance and renewals for apartment operators and patient paperwork for specialty physician groups, raised $350 million at a $4 billion valuation. That is double its value in August 2025. Andreessen Horowitz and Bessemer led the round. The company says one in six US apartments uses its software, and its annual recurring revenue passed $200 million this summer. It is a clear example of AI that sells because it takes over one industry's routine admin work.

Physical AI

Humanoid robots are shipping in real numbers, and they are cheaper than many people assume. IDC counts about 25,000 humanoid units shipped worldwide in the first half of 2026, up 432% on the year, for a market worth $740 million, up 323%. That works out to under $30,000 per unit on average. Agibot passed Unitree as the largest maker, with more than 8,600 units and 35% of global share. Unitree grew 170% to about 5,900. China accounts for about 78% of shipments by IDC's count. The mix is also changing. Research, education and demonstration uses fell from 84% of shipments in 2025 to 69%. Manufacturing, logistics and warehousing now make up most of the rest. For a US buyer, that means supply, prices and tariff exposure will mostly run through Chinese manufacturers.

Tesla shows the gap between building robots and making them useful. The Information reports that Tesla could make about 1,000 Optimus humanoids a week by year-end, ten times its second-quarter pace. But the robots handle tasks they were trained on and behave unpredictably on anything new. Even a simple new job can take days to teach. The hands, with more than 100 parts, are hard to assemble without rework, and the touch sensors wear out. Tesla's fix is a replaceable sensor "glove." A separate report from DigiTimes says Tesla is turning to Chinese manufacturing to support the Optimus supply chain ramp. That is the same approach it took with electric vehicles.

Investor Nikunj Kothari put the numbers side by side in an essay titled "Deployment Gap." Humanoid manufacturing costs fell about 40% in a year. One maker's robots have logged more than 65,000 hours across nine customer sites, and a 16-robot fleet reached a 95% average success rate. But carmakers are piloting "single or double digit numbers of robots per plant," and robots shown off in public are often not on actual production lines. Robotics startups raised $18.8 billion in the first half of 2026, more than in all of 2025. Kothari's advice to investors is also good advice for buyers: visit the plant and talk to the plant manager.

New designs are aimed at cost rather than show. Korea's Neubility unveiled Billy, a wheeled humanoid with two arms and a torso that adjusts in height. It navigates by camera instead of lidar. The cameras cost $50 to $500, against $1,000 to $10,000 for lidar, though camera navigation struggles in low light and in repetitive spaces. Trials start in the second half of 2026 in food-and-beverage plants, logistics and transport hubs. The company has not announced a price. Neubility builds on nine years of field data from more than 44,000 deployments of its Neubie delivery robot.

Quick Takes

  • America.gov. The White House launched an AI-powered site meant to be a single front door to federal services, from passports to Medicare enrollment. It is overseen by Chief Design Officer Joe Gebbia, and an executive order tells agencies to connect their services to it.

  • Walmart bans AI-made store signs. Stores can no longer use ChatGPT, Claude or Gemini to create displays and promotional graphics. Signage must come from the corporate catalogue, closing a loophole after customers reacted badly to AI visuals.

  • Codex Security Cloud and Decisions API. Two quieter DevDay launches. One scans code repositories for vulnerabilities on a schedule. The other returns a fixed answer from a set you define, which fits routing and classification jobs.

What This Means for Your Business

Pilot an always-on agent on one job with a clear finish line, and keep yourself as the final approver. Dots and Meta's Muse can now watch your accounts and act. Start with something like chasing unpaid invoices or flagging unanswered customer emails. Give read-only access first, require your sign-off before anything is sent or paid, and check a week of its work before you widen its permissions. OpenAI's invoice example works because the human approved the send. Keep that step.

Budget for AI as a utility, not a seat. Microsoft now meters Copilot's agent features, OpenAI's top plan is $500 a month, and agents run when no one is watching. Before you turn on any agent feature, set organization and per-user spending caps, switch on usage alerts, and name one person to review the bill monthly. If a vendor can't show you per-agent usage, treat that as a warning sign.

Re-price your AI workloads. GPT-6.1 Sol and Claude Sonnet 5.5 both list at $2/$10 and come close to flagship results on many tasks. If you are paying flagship rates for drafting, summarizing or routine coding, test the mid-tier model on a week of real work. You may cut that line by as much as 80% with little loss. Keep the expensive model for jobs where a mistake is costly.

Tighten security around your agents' connections. The MCP toolkit flaw and Anthropic's GLM-5.3 findings point the same way: attacks are getting cheaper, and agent logins are a new way in. Ask every vendor and developer who built you an AI connector which MCP SDK version they use. Give each agent its own credentials with the narrowest permissions possible, and revoke tokens you don't recognize.

On robots, check the reference site before the demo. Humanoids are shipping in the tens of thousands at an average under $30,000, but even Tesla's take days to learn a new task. If a vendor pitches you, ask for uptime and success rates from a live site running the same job you have, who retrains the robot when your process changes, and where the hardware is made. Most of it is made in China, so ask about tariff exposure too.

Sources

← All posts