Agents of Work
Let's Talk
October 8, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 8, 2026

Summary

The free version of the most widely used AI tool just got a lot more capable, and the cheap end of the market got cheaper still. ChatGPT now answers with working buttons, charts and small tools for every user, Anthropic cut the price of its smallest model by about three quarters, and Windows is being rebuilt as a place where AI agents run on the desk instead of in the cloud.

Highlights

  • Free ChatGPT users now get GPT-6, so your staff and customers have a stronger assistant at no cost.

  • ChatGPT can now build calculators and checklists inside a chat; simple calculator pages on your site may lose traffic.

  • Claude Haiku 5.5 costs about 75% less than Haiku 4.5; reprice any high-volume AI job you run.

  • Claude Max and Team subscribers get monthly API credits, up to $500 a month for Team plans.

  • Fake ChatGPT and Claude ad portals are stealing ad-account logins; switch ad accounts to phishing-resistant sign-in.

  • Google's SynthID Detector is now free for anyone to check images, video and audio from major AI tools.

  • A $2,599.99 Surface laptop can run AI models of up to 120 billion parameters without the cloud.

  • A $92,478 stake in a Chinese robotics firm drew a $200,000 US fine; check where your funds invest.

  • Zalando's returns are now sorted by robots rented as a service, not bought outright.

Quick Takes

  • Grok Bot will borrow its rivals' models. Elon Musk said on X that Grok Bot, the agent app from SpaceXAI, will no longer run only on SpaceX's own models and will route tasks to Anthropic's Claude Opus 5.5, Midjourney, Suno and "other leading APIs," picking whatever "is most likely to give you the best outcome." He did not say which tasks go where. The move came a day after users reported trouble reaching Grok on mobile and the web. It is another sign that agent apps are becoming brokers that shop among models rather than showcases for one. Read more

  • Anyone can now check for AI watermarks. Google opened its SynthID Detector to everyone, globally and in English, at synthid.com. It checks images, video and audio for watermarks from Google tools and from partners OpenAI, NVIDIA and Kakao, with Apple listed as coming soon. Google says it has watermarked more than 180 billion images and videos. The catch is that it only finds content made with tools that add SynthID, so a negative result does not prove a file is real. Read more

  • AI mortgage servicing raises $150 million. Valon, which builds software and AI to run mortgage servicing, raised a $150 million Series D at a $2.3 billion valuation. Investors are paying up for AI that runs whole back-office operations in regulated industries, not just chat on top of them. Read more

What's Covered in Featured News

  • GPT-6 reaches every ChatGPT user: which model each plan gets, what Intelligent UI builds inside a chat, and what it means for websites.

  • Claude Haiku 5.5: Anthropic's new small model, its price cut, the business test results and the new monthly credits.

  • Windows as an agent computer: NVIDIA's RTX Spark chips, Microsoft's Execution Containers and the Surface Laptop Ultra.

  • Phishing dressed as AI ad tools: how fake ChatGPT, Gemini and Claude portals steal ad-account logins and codes.

  • OpenAI's math drop: hundreds of claimed results, a proof of the Unique Games Conjecture and the open questions.

  • Texas makes data centers wait: why the state froze permits and what its new fees and power rules require.

  • Physical AI: Sereact's returns robots at Zalando, Beam Global's US-built drones, the first US fine on outbound robotics investment, Amazon's robot toolchain, a new Boston Dynamics chief and a new safety standard for robot users.

Featured News

GPT-6 reaches every ChatGPT user, and answers become small apps

OpenAI began moving all of ChatGPT to GPT-6 this week. On October 7, Plus, Pro, Business and Enterprise users got GPT-6 Sol in the Chat tab; on October 8 the rollout reached the Free and Go tiers, which get GPT-6 Luna, the cheaper model OpenAI tuned for speed. Sol and Luna launched on September 22 for ChatGPT Work, Codex and the developer platform but had not been in everyday chat until now. The top-end Astra model still powers the Pro thinking level. Enterprise access depends on admin settings, and the models behind ChatGPT Work and Codex are unchanged.

The bigger change is what an answer looks like. OpenAI calls it Intelligent UI: instead of paragraphs, ChatGPT can reply with buttons, forms, charts, diagrams, maps and checklists, and the model decides the mix based on the question. In OpenAI's own side-by-side examples, a recipe answer becomes a widget that rescales ingredients for the number of guests, and a request to split a bill produces a working bill splitter. Users can also ask it to build a simple tool for the job at hand, such as a calculator. Under the hood it draws from a library of native components that stream in while the model writes. There is no switch to turn it on, though on the web a "Layout and visuals" setting can turn most of it off. It works on the web and in updated apps but not in older ChatGPT desktop apps for Mac and Windows, and Astra does not support it.

OpenAI also says GPT-6 Instant starts answering web-search questions 44% sooner on average than GPT-5.6 Instant, partly because it can begin replying before it finishes thinking or using tools. That figure comes from OpenAI's internal tests.

For a business, two effects follow. Inside the company, every employee with a free account now has a stronger assistant that can produce a working checklist or quote calculator on request, which makes it more important to say what data staff may paste into it. Outside, the simple tools many small firms built to attract visitors, such as mortgage, savings and unit calculators, are exactly what ChatGPT can now produce inside the chat, so customers may never need to click through.

Anthropic cuts the price of its small model by about three quarters

Anthropic released Claude Haiku 5.5 on October 7 and priced it to win high-volume work. For requests up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, against $1 and $5 for Haiku 4.5. Anthropic says requests of that size are 90% cheaper, larger ones 50% cheaper, and average costs about 75% lower, after accounting for a new way of counting text that uses slightly more tokens per task. It is aimed at summaries, sorting and classifying, database lookups, live customer support and browser tasks, and at serving as a fast helper to the larger Opus 5.5 and Sonnet 5.5 models. For the first time in the Haiku line, users can dial effort up or down to trade cost against quality.

The early business results are concrete. HubSpot said Haiku 5.5 scored 92.8% on its CRM test suite, the best it has seen from a smaller model, and Box said it scored 11 points higher than Haiku 4.5 at about half the response time. Anthropic also halved cache-read prices on Sonnet 5.5 and is giving subscribers monthly API credits for building agents: $100 on Max 5x, $200 on Max 20x and up to $500 pooled across a Team plan. The model is available on AWS, Google Cloud, Microsoft Azure and Anthropic's own platform.

Windows is being rebuilt as a computer for agents

At a Microsoft event in San Francisco, Jensen Huang and Satya Nadella presented a joint plan to run AI agents locally on Windows PCs. Microsoft made Execution Containers generally available: operating-system-level boxes that let an agent run in the background while a company decides which files and networks it can touch. NVIDIA's new RTX Spark chips pair a Blackwell graphics processor with up to a 20-core Grace processor, and laptops using them from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte went on preorder October 7, with availability from October 16 and compact desktops in November. Microsoft's Surface Laptop Ultra starts at $2,599.99 with up to 128GB of shared memory, and Microsoft says it can run local models of up to 120 billion parameters. For firms that handle client files they would rather not send to a cloud service, a capable local AI machine now costs about as much as a high-end laptop.

Phishing dressed up as AI advertising tools

Researchers at the browser-security firm Island found a phishing platform that impersonates advertising products from ChatGPT, Gemini, Claude, Perplexity, Meta's Muse and Manus. Victims, mostly agency staff, media buyers and ad-account administrators, receive invitation emails, click "Connect" and see a fake sign-in window drawn inside the page with a fake address bar showing a real Google or Okta address. A human operator watches in real time, chooses which verification prompt to show next, collects the code and signs in to the real account. The likely aim is to run the attackers' own campaigns on accounts with clean spending histories. The same infrastructure also runs fake Google Ads refund notices and fake job sites. The researchers advise phishing-resistant sign-in, such as passkeys or hardware keys, and caution before connecting any AI tool to an ad account.

OpenAI's math drop and what it does not settle

OpenAI released hundreds of mathematical results produced by an unreleased internal model. The computer scientist Scott Aaronson, writing on October 7, reported that the model attempted about 8,000 problems and solved roughly 5% of them, using about three hours of top-tier compute per solved problem. Among the claims is a proof of the Unique Games Conjecture, a major open question in computer science, accompanied by a machine-checked certificate in the Lean proof language. Not every result carries one, and Aaronson notes that most of the arguments have not yet been fully understood by a person. An outside advisory group said only mathematicians can judge the work. For business readers, the takeaway is that AI is producing work faster than experts can check it, and checking is where the bottleneck now sits.

Texas makes data centers wait

Texas has moved from asking data centers to pay for their own grid upgrades to freezing new data center permits, a shift Ryan McEntush, an investing partner at Andreessen Horowitz, laid out in detail on October 7. Requests from large new customers on the state grid, run by ERCOT, grew from 63 gigawatts at the end of 2024 to 474 gigawatts by June 2026, more than five times record peak demand, and about 90% of it is data centers. Many are speculative filings on several sites at once. New state rules effective October 8 add a $100,000 study fee and a $50,000-per-megawatt deposit, and offer faster paths to campuses that bring their own power or accept automatic cuts when lines are full. Grid strain is now a direct brake on how fast AI capacity grows, and on electricity prices for everyone sharing the grid.

Physical AI

Robots are taking on one of retail's messiest jobs. Sereact's dual-arm robots are now sorting returned clothing for Zalando at two CEVA Logistics sites, in Greven, Germany, and Świebodzin, Poland. Its software identifies, grasps and sorts items it has never seen before, without training for each product, which matters because returns arrive crumpled and badly packed. "Returns are the hardest manipulation problem in fashion logistics," said Sereact chief executive Ralf Gulde. CEVA offers the system as Robotics-as-a-Service, so the warehouse rents the capability rather than buying it. The companies did not disclose the number of robots, throughput or costs, which is what an operator would need to judge the payoff.

Where a robot is made is becoming part of its price. Beam Global agreed to buy ScoutDI, a Norwegian maker of drones that inspect tanks, cargo holds and other confined spaces, for about $24 million in cash and stock, with closing expected in November. Beam plans to build ScoutDI's systems for the US market in its American factories. ScoutDI's Scout 137 Gen3 has conditional approval from the US Department of Defense, which keeps it off the FCC's Covered List of restricted equipment as long as the company follows through on moving production to the US. Its customers include ExxonMobil, Chevron and Ørsted. In a related signal, the US Treasury fined Amidi, the parent of the Plug and Play Tech Center accelerator, $200,000 for failing to report a roughly $92,478 investment its Chinese fund made in Noematrix, a Shanghai robotics and AI company. It is the first penalty under the outbound investment rules in force since January 2025, and it is more than twice the size of the stake.

The tools for building robots keep consolidating. Amazon launched an open-source Physical AI Toolchain on AWS that strings together synthetic data, training, simulation, deployment to machines and fleet updates, built on NVIDIA's Isaac and Cosmos software. NEURA Robotics is an early user, and Amazon did not list pricing. Boston Dynamics named Rohit Prasad, who helped build Alexa and led Amazon's Nova models, as chief executive effective October 7, filling a post vacant since Robert Playter stepped down in February.

For operators already running mobile robots, the industry body A3 published ANSI/A3 R15.08-3-2026, the final part of its industrial mobile robot safety series. This part is written for the companies that use the robots day to day, covering risk assessment, managing changes, staff training and periodic inspections. It is sold as a PDF from A3's online store.

What This Means for Your Business

Update your AI use policy this week. With GPT-6 now on free ChatGPT accounts, more staff will use it for real work whether or not you pay for a business plan. Personal accounts sit outside whatever admin controls your company has set up. Write down which kinds of information may never go into a personal account, such as customer records, contracts and pricing, and consider whether a paid business plan is cheaper than the risk.

If your website relies on simple tools to draw visitors, such as a quote calculator, a sizing guide or a savings estimator, assume ChatGPT can now produce a version of it inside the chat. Check your analytics for those pages over the next two months. The defense is a tool that uses information only you have, such as your actual prices, stock or delivery times, which a general assistant cannot copy.

If you run any high-volume AI work, such as tagging support tickets, summarizing calls or sorting invoices, reprice it now. Haiku 5.5 is roughly a tenth of Haiku 4.5's price for typical requests. A job that was not worth automating at last year's prices may be now. If you pay for Claude Team or Max, make sure someone claims the monthly API credits before they go to waste.

Lock down ad accounts. Anyone who manages your Google, Meta or other advertising accounts should switch to passkeys or hardware security keys, and treat any email inviting them to "connect" a new AI ad tool as suspicious until verified through the vendor's own website. Review who has admin rights on each ad account and remove anyone who no longer needs them.

If you buy drones, robots or other connected equipment, ask where it is made and whether it appears on, or is exempt from, the FCC's Covered List. The Beam deal and the Treasury fine show that country of origin now affects both what you can legally deploy and what investments are allowed. And if you rent warehouse automation, ask for throughput and uptime figures in the contract, not just a demo.

Sources

← All posts