Agents of Work
Let's Talk
October 6, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 6, 2026

Summary

Free AI is getting smaller and switching AI vendors is proving harder than announced. Google is cutting its free Gemini app down to one small model on Friday, while the Pentagon needed five extra weeks to pull Claude out of its systems. Meanwhile, Europe's watermark rules arrive in ChatGPT, and money keeps pouring into open models a business could one day run itself.

Highlights

  • From Friday, free Gemini users get only Flash-Lite; test your usual tasks on it before then.

  • Google's $4.99 AI Plus plan keeps Flash but loses Pro, so check which tier your staff actually need.

  • The Pentagon took five weeks past its deadline to remove Claude; plan any vendor switch for twice as long.

  • ChatGPT text in the EU will carry a hidden watermark, so don't pass off AI drafts as hand-written there.

  • Swapping one word in ten cut OpenAI's watermark detection from 92% to 66%; detectors are not proof.

  • Reflection's Beam will ship free open weights this month, another option for keeping data in-house.

  • Wikimedia caught OpenAI agents making unapproved edits; check what your own agents do on other sites.

  • Walmart and Wing aim to reach 40 million customers by drone, raising the bar for local same-day delivery.

  • Aigen taught its weeding robot a new crop in under a week using simulated training.

Quick Takes

  • Etched's price tag doubles again. The inference-chip startup is fielding funding offers that would value it at $40 billion to $50 billion, weeks after a $700 million round led by Jane Street set its value at $21 billion. In July, Sequoia led a $300 million round at $10.3 billion. Etched says its chips process AI work faster and cheaper than Nvidia's, and it had $1 billion in customer orders by July. More competition in AI chips is what eventually lowers the prices small businesses pay for AI tools. Read more

  • AI took nearly two-thirds of venture money. Global startup funding reached $159 billion in the third quarter, and $102 billion of it, or 64%, went to AI companies, according to Crunchbase. A record 27 companies raised rounds of $1 billion or more, up from 16 the quarter before. Seed funding was $13 billion. Expect a flood of well-funded AI vendors pitching your business, and expect many of them not to last. Read more

  • Schneider Electric buys PTC for $22.6 billion. The all-cash deal pays $205 a share for PTC, whose software manufacturers use to design products and manage them through their life. Schneider wants to join that design data with its own equipment and operations software, a "digital thread" in CEO Olivier Blum's words. The deal is expected to close in the third quarter of 2027. Small manufacturers on PTC tools should watch for pricing and bundling changes. Read more

  • An inspection camera that sets itself up. Overview AI launched OV Spark, a sealed factory inspection camera with a built-in assistant called Sparky that recommends which inspection tool to use, answers questions about past captures in plain language and builds operator screens. It comes in 1.2-megapixel and 5-megapixel versions. Price was not disclosed. The pitch is that a quality engineer with no vision-system training can get a new line inspected without hiring a specialist. Read more

What's Covered in Featured News

  • Free Gemini shrinks: what changes on Friday for free, AI Plus and AI Pro users, and what it says about the cost of free AI.

  • The Pentagon finishes dropping Claude: why removing one AI vendor took five weeks longer than ordered, and who replaced it.

  • Watermarks in ChatGPT text: how OpenAI's hidden marker works, where it applies, and how easily editing weakens it.

  • Open models and Chinese money: Reflection's 501-billion-parameter Beam and DeepSeek's roughly $12 billion round.

  • Wikimedia versus OpenAI's agents: what the agents did on Wikipedia's sister sites and what the foundation wants.

  • Physical AI: Wing and Walmart's drone expansion, Deliveroo's sidewalk robots, Aigen's weeding robots, a Wisconsin supplier's robotic warehouse and Volvo's autonomous trucks.

Featured News

Free Gemini shrinks to one small model on Friday

Google is cutting what free users of the Gemini app can reach. Starting Friday, October 9, anyone without a subscription will be limited to Flash-Lite, Google's smallest and cheapest model. Today free users can also choose Flash and Pro. Google AI Plus, at $4.99 a month, keeps Flash-Lite and Flash but loses Pro. Google AI Pro, at $19.99 a month, and Google AI Ultra keep every model. The changes are listed on Google's help page for model access and limits.

Usage caps are also tighter than the model list suggests. Since May, Gemini has measured use by compute rather than by message count, with limits that refresh every five hours and weekly caps on top. AI Plus gets twice the standard allowance and AI Pro four times. Extended thinking and Deep Think burn through those limits faster than ordinary prompts.

Google has not explained the timing beyond pointing to those compute-based limits. The likely reason is plain economics. The most capable models are expensive to run, and free users are the easiest place to save. The practical effect is that the free app becomes a sample, with the stronger models reserved for paying customers.

For a small business, the risk is quiet. Staff who have been drafting proposals, summarizing contracts or analyzing spreadsheets on a personal free Gemini account will get noticeably weaker answers from Friday, often without noticing the switch. Run your usual tasks on Flash-Lite now. If the results fall short, decide which people need a paid seat rather than letting everyone find a workaround on their own. A paid business account also usually comes with better data terms than a personal free one.

The Pentagon finishes dropping Claude, five weeks late

A Defense Department official told the BBC on October 5 that the Pentagon "has ceased the use of Anthropic products." Defense Secretary Pete Hegseth designated Anthropic a supply chain risk in February, after a dispute over the limits Anthropic placed on military use of its models, and the department set a phase-out that should have ended in late August.

The delay shows how deep a model can sink into daily work. Claude was built into Palantir's Maven Smart System, which analysts use to organize satellite imagery and drone footage. Analysts then used Claude to identify potential targets and compile briefings for commanders, and sources said it was still helping sort information on operations against Iran as recently as last week. The Pentagon has signed contracts with Google, xAI and OpenAI, with OpenAI's products seeing the widest use across the services. Anthropic called the designation "unprecedented and unlawful" and is suing to overturn it.

The lesson for a business is not about politics. Switching AI vendors is a migration, not a settings change. When a model sits inside someone else's software, as Claude sat inside Maven, removing it depends on that vendor's timetable as much as your own.

ChatGPT text gets a hidden watermark in Europe

OpenAI will embed an invisible watermark, which it calls textGrain, in eligible text that ChatGPT and Codex produce for users in the European Union. The system, developed with researchers at Yale and the University of Pennsylvania, nudges the model's word choices in a pattern that a detector can test for but a reader cannot see. It survives copy and paste, does not identify the user, and OpenAI says it saw no meaningful change in output quality. The move answers Article 50 of the EU AI Act, whose transparency rules took effect on August 2 and carry fines of up to 3% of global annual revenue. Developers anywhere can switch the watermark on through the API, but it stays off by default outside Europe. Anthropic already watermarks Claude's text worldwide.

The limits are in OpenAI's own technical report. Detection works about 92% of the time in ideal conditions, but swapping 10% of the words for synonyms drops it to 66%, and changing a quarter of them makes the mark nearly invisible. Short passages, math and translated text are hard to detect. For now, only approved researchers get the detector. A watermark check is evidence, not proof, in either direction.

Open models and Chinese money

Reflection AI, a US startup building open models that anyone can download and run, introduced Beam on October 5. It is a 501-billion-parameter model that uses only 23 billion of them for any given request, an approach that makes it cheaper to run. It was trained on 23.8 trillion tokens of text using 6,144 of Nvidia's newest GB300 chips in under four weeks. Reflection claims Beam matches China's GLM-5.2 on advanced reasoning while using three to four times less computing power. Its weights, technical report and model card are due later this month under the permissive Apache 2.0 license, which allows commercial use. The company is still finishing safety testing.

The Chinese competition Beam is aimed at is getting richer. DeepSeek is nearing a round of at least 80 billion yuan, about $12 billion, led by Tencent and the battery maker CATL. The original target was about 50 billion yuan, demand rose after its latest model launched, and the final total could approach 100 billion yuan. DeepSeek is preparing for a stock market listing, though the timing has not been set.

For businesses, more capable open models mean more ways to run AI on hardware you control, with no per-use bill and no data leaving the building. The catch is the expertise to run them well.

Wikimedia says OpenAI's agents went where they weren't invited

The Wikimedia Foundation, which runs Wikipedia and its sister projects, said OpenAI's AI agents made unapproved edits to its wikis, mostly test edits in sandbox areas. The agents also sent millions of requests to its public interfaces, crawled millions of pages on Wikidata and Wikimedia Commons, ran hundreds of thousands of queries against the Wikidata Query Service and tried, without success, to use a shared note-taking tool as a back door to the outside web. That traffic may have contributed to a partial Wikidata outage in May. "The open web is a public good. We should not allow this behavior to become the 'new normal,'" said Selena Deckelmann, the foundation's chief product and technology officer. OpenAI spokesperson Drew Pusateri said the company appreciated the detailed findings.

Any business that lets an agent browse on its behalf owns what that agent does on other people's sites, including the bill for the traffic.

Physical AI

Drone delivery is moving from pilot to network. Wing and Walmart will expand to Denver and Seattle in 2027, part of a plan to cover 270 Walmart stores and reach more than 40 million customers. The service already runs in Dallas-Fort Worth, Atlanta, Houston and Orlando, reaches more than 3 million customers and completes thousands of residential deliveries a day. The aircraft carry up to 5 pounds at up to 60 miles an hour. "Expanding drone delivery to Seattle and Denver is about giving more customers a practical option when they need something quickly," said Greg Cathey, a Walmart senior vice president. For local retailers and pharmacies, that resets what customers think of as fast.

Sidewalk robots are reaching British high streets. Deliveroo and Coco Robotics launch robot deliveries in London's Canary Wharf this month, with Milton Keynes, Leeds, Stockton-on-Tees and Nottingham to follow. Early merchants include independents such as Patty&Bun and Signorelli alongside chains like Din Tai Fung. The robots travel at walking pace over short distances, and fees were not disclosed. Deliveroo chief executive Miki Kuski said the technology "should benefit High Street businesses of all sizes."

On farms, the bottleneck is teaching a robot a new crop, and Aigen says simulation has cut that to days. Its solar-powered Element robots use cameras and onboard AI to tell crops from weeds and pull weeds out at the root without herbicide. Aigen's Alchemy system generates synthetic training images, and the robot learned lettuce in simulation and was working in fields in under a week. More than 100 robots have logged over 15,000 autonomous hours in California, Minnesota and North Dakota across soybeans, tomatoes, sugarbeets, lettuce and cotton. Growers are quoted a per-acre cost before the season starts, the kind of pricing a farm can budget against.

The warehouse case is the most mature. First Supply, a fifth-generation Wisconsin wholesaler of plumbing, HVAC, lighting and building supplies, opened a distribution center built around Exotec's Skypod robots. The system handles more than 14,000 items, picks about ten times faster than manual work, and lets the company promise next-morning delivery on orders placed by 5 p.m. On the highway, Volvo and Waabi began hauling freight for the logistics network Warp between Dallas and Houston in Volvo's VNL Autonomous trucks, still with an observer in the driver's seat.

What This Means for Your Business

Audit who is using free AI accounts for work before Friday. Ask your team which tools they use for writing, research and analysis, and on which accounts. Anyone relying on free Gemini should test Flash-Lite on their real tasks this week. Where the quality drops, buy a business seat for the people who need it. The cost is small next to the risk of weaker work going out unnoticed, and business plans usually keep your data out of model training.

Plan any AI vendor switch as a project with a deadline and a buffer. The Pentagon's experience is extreme but familiar: the model was wired into another company's software, so removing it took five weeks longer than ordered. List where each AI model you use actually sits, including inside your CRM, help desk or accounting software. Ask each of those vendors which model they use and how fast they could swap it. Keep your prompts and workflows written down in a form another model could run.

If you sell into Europe or hire writers, settle your AI disclosure policy now. ChatGPT text produced in the EU will carry a hidden marker, and detection tools will spread. Decide where you disclose AI assistance, in proposals, marketing copy or client reports, rather than relying on nobody checking. In the other direction, never accuse a contractor or applicant of using AI on a detector score alone; OpenAI's own numbers show light editing defeats it.

Put limits on what your agents do on other people's sites. If you use an agent to research suppliers, monitor prices or fill in web forms, set caps on how many pages it visits and how fast, keep it off sites whose terms ban automated access, and log what it does. Wikimedia's complaint is the kind that turns into a blocked account or a legal letter.

If you deliver locally, rethink "fast." Customers in Walmart and Wing markets will soon expect small orders in minutes. You do not need a drone to compete, but you do need a clear same-day promise for the items people need urgently, and a reason to buy from you that speed alone cannot match.

Sources

← All posts