Agents of Work
Let's Talk
November 3, 2026 · Agents of Work

Agents of Work AI Daily Briefing — November 3, 2026

Summary

The companies selling AI are showing their bills, and the costs are starting to land on customers. Anthropic's amended IPO filing reveals $895 billion in compute commitments, Grok and ChatGPT are reworking usage allowances, and a new test finds the best model finishes only 12% of real expert work, so the gap between what AI costs and what it reliably delivers is today's story.

Highlights

  • Anthropic made about $17 billion in Q3, but its quarter-over-quarter growth slowed to roughly 47%.

  • Anthropic has committed about $895 billion to compute, so expect AI vendors to keep pushing prices and limits.

  • The best AI finished 12% of realistic expert tasks, so treat agents as drafters, not finished-work machines.

  • Every model scored zero on pass-fail legal work in the same test; keep lawyers on contracts.

  • WorkOS says attackers used AI to automate a breach, then aimed phishing at its customers with real conversation details.

  • Verify any unexpected payment or data request by phone, even when it quotes a real thread.

  • Proton dropped plans to re-interview staff about AI use after backlash; measure results, not hours in the tool.

  • ChatGPT and Codex will share one allowance, so heavy coding use may squeeze everyday chat.

  • A robot lamp that folds laundry starts at $2,699 on preorder, a home robot price point worth noting.

Quick Takes

  • Grok Bot users hit their limits early. Some subscribers saw their weekly usage jump from about half used to maxed out over Halloween weekend, and Elon Musk confirmed on X that it was deliberate, writing that "the cost of serving Grok Bot was unsustainably high relative to revenue." He has not said what the limits will be going forward. If a workflow depends on one AI tool's allowance, have a fallback. Read more

  • Squarespace cut about 300 jobs, roughly 19% of its staff. The company calls it a restructuring to "move faster" and did not blame AI, but the cuts land as prompt-to-website builders crowd its market and as Squarespace prepares Forma, an AI-native platform. If your website runs on Squarespace, nothing changes today, but watch for product and pricing shifts. Read more

  • Claude's hidden watermark now covers older models too. Anthropic is extending its invisible text watermark, built in response to EU rules, to models including Opus 4.7, Opus 4.6 and Sonnet 4.6, worldwide and with no opt-out. Only eligible groups such as regulators, journalists and researchers can detect it, and Anthropic concedes short texts may not carry enough of the pattern. If you publish Claude-drafted copy, assume it may be identifiable later. Read more

  • Claude agents get faster on long jobs. Anthropic's new message threads feature, in public beta, lets an agent send only new messages instead of the whole conversation each turn. For very long sessions, median time to the first word of a reply drops about 35%, though usage and pricing are unchanged. Read more

What's Covered in Featured News

  • Anthropic's amended IPO filing: the third-quarter numbers, the $895 billion compute bill, and the warning about AI that helps build AI.

  • The GDPx test: what happened when top AI models were handed eight-hour expert tasks in law, finance, chip design and more.

  • ChatGPT merges Chat and Work: what changes in the desktop app and what the shared ChatGPT–Codex allowance means.

  • An AI-assisted breach at WorkOS: what was taken, and why the follow-up phishing is the real risk for customers.

  • Proton's AI re-interviews: why a CEO dropped plans to grill staff on their AI use, and what it says about measuring adoption.

  • Gartner on AI in IT operations: where agents are headed by 2030 and why usage-based costs are the catch.

  • Physical AI: Apple's rumored robotic HomeView, Syncere's $2,699 laundry-folding lamp, GenRobot's $100 million for robot training data and NVIDIA's new Cosmos tools.

Featured News

Anthropic's filing shows $17 billion a quarter, and an $895 billion bill

Anthropic released an amended IPO prospectus on November 2 that adds third-quarter results to the first-half numbers it filed the week before. Revenue for the nine months through September 30 was $33.24 billion, up roughly 1,200% from a year earlier. Third-quarter revenue alone was about $17 billion. Growth is still enormous, but it is slowing: revenue rose about 47% from the second quarter to the third, against 144% in the second quarter and 132% in the first.

The losses are just as large. Anthropic reported a $104 billion net loss for the nine months, mostly from changes in the fair value of convertible notes and derivative liabilities rather than from operations, and a $23.29 billion loss in the third quarter. Compute and infrastructure expenses rose $14.4 billion over the nine months, a 329% increase. The figure that will shape the industry is in the commitments section: purchase commitments and leases total about $895.4 billion, "substantially all of which is related to compute and infrastructure." Much of that capacity comes from Amazon, Google and SpaceX. No IPO date or valuation has been set, and CEO Dario Amodei and his co-founders included more than 90 pages of risk factors.

One of those risk factors drew as much attention as the money. The amended filing says Anthropic's internal estimates "indicate that we may imminently meet" its own policy's threshold for "dramatically accelerated progress attributable to AI," meaning AI meaningfully speeding up the building of the next AI. Anthropic says no current model has crossed that line, that it has in the past spotted such accelerations only after they happened, and that it does not expect to safely let Claude build future models without "meaningful human supervision or understanding." Skeptics note an S-1 is also a sales document.

For a business owner, the takeaway is the cost structure. A vendor that has signed for nearly $900 billion in compute needs customers to pay more over time, and this week's usage-limit changes at xAI and OpenAI point the same way. Budget for AI as a rising line item, not a falling one.

The best AI finished 12% of real expert work

SemiAnalysis published GDPx on November 2, a test that drops AI agents into realistic professional work across 11 fields, including chip design, cybersecurity, biology, finance, law and materials science. Tasks come with real material, such as a deal room from a private equity transaction, and most allow up to eight hours. Even the strongest model scored just 12% overall. Anthropic's Claude Opus 5.5 was the best all-rounder, Google's Gemini 4 was strongest in biology and OpenAI's GPT-6 Astra led in materials science.

Finance and law were the hardest. On a legal redlining task, every model scored zero under pass-or-fail grading, so the authors added partial credit: Opus 5.5 averaged 44% and GPT-6 Astra 25%. No model fully solved the finance task either; the best single run passed 8 of 11 checks. Two caveats matter. SemiAnalysis designed and grades the test itself with outside partner firms, and safety filters blocked some models from cyber and biology tasks, pulling their scores down.

The practical reading is that today's best models can work for hours on expert material and still rarely finish the job unaided. They are strong first-draft engines, and the review step is where the value gets protected.

ChatGPT folds Work into Chat

OpenAI emailed desktop app users that the separate Chat and Work tabs are being merged into one experience, starting with a test group before a wider release. In the combined view, users can bring in files and plugins and ask ChatGPT to build a website, create a presentation or carry out web tasks while they review its output. Codex stays a separate experience in the app. ChatGPT and Codex will also draw from a single usage pool, and OpenAI says that while most people won't notice, some may have to wait for limit resets more often. Anthropic made a similar move in September, folding Claude's chat and Cowork features into one interface. For teams, the simpler interface is welcome, but whoever writes code in Codex may now be spending the same allowance others use for everyday writing.

An AI-assisted breach, and the phishing that follows

WorkOS, which sells login and identity tools to software companies, disclosed that on October 30 it found unauthorized access to an internal application used by its sales and customer-facing teams. Exposed data included business contact details, account histories, operational data about how customers use WorkOS, and summaries and excerpts of sales calls, support discussions and customer Slack conversations. The company says its investigation found an AI-assisted offensive security framework automating reconnaissance and exploitation attempts, and that there is no evidence its production services or application data were touched.

The more useful part of the bulletin is the warning. Because attackers now hold real names and real conversation history, WorkOS tells customers to verify unexpected requests before acting, even when they reference a real colleague or support thread. That advice applies to any small business whose vendors keep notes on them, which is all of them.

Proton backs off grading staff on AI use

Proton, the privacy company behind Proton Mail and Proton VPN, canceled plans for nearly company-wide "re-interviews" in which employees would explain how they use AI. According to six current employees and audio from a roughly two-hour all-hands, staff understood the interviews as possible grounds for firing people who weren't using AI enough. CEO Andy Yen dropped the plan the day after an investigation into his push for AI, which included an August exchange in which an employee asked for a few days of security testing before releasing code. Yen called it "loser thinking" and told staff to "stop lovingly handcrafting all our code and use AI more." Users called for his resignation.

The lesson for any manager is that counting AI use measures activity, not results, and can make the careful employee look like the slow one. Ask what got shipped and what got checked.

Gartner: agents are coming to IT operations, with a cost catch

A Gartner survey of 272 heads of IT infrastructure and operations found about 58% expect scaling AI pilots to be a priority within 12 months. Virtually no organizations run AI agents with minimal human involvement in IT operations today, but Gartner expects half to by 2030, handling tasks like incident triage, change validation and fixes. Gartner VP analyst Jonathan Forest urged leaders to focus on "a small number of strategic investments," and the firm warned that rising usage could bring costs nobody budgeted for. For a small IT team or managed service provider, that is the same lesson as Anthropic's filing: price agents by what they will use, not what the demo used.

Physical AI

Apple's home robot is getting more concrete, if not yet official. Bloomberg's Mark Gurman reports that the next version of Apple's HomeView smart display, codenamed J595, would add a motorized arm that lets a 9-inch screen turn 360 degrees to face whoever is speaking, up from the 7-inch screen on the $299 HomeView that exists now. The report points to a 2027 or 2028 release and names no price; a $1,000 figure circulating again this week dates to 2024 reporting and is unconfirmed. Apple has not commented. For businesses, a FaceTime display that tracks the person talking is a conference-room and front-desk product as much as a home gadget, but it is at least a year away.

Home robots that do chores are arriving at a price people can actually compare. Syncere, a Palo Alto startup founded less than a year ago by robotics researchers Aaron Tan and Angus Fung, raised an $11 million pre-seed round led by a16z Speedrun. Its product, Lume, is a floor lamp with a mobile robot built in that is designed to fold laundry, tidy rooms and make beds, starting at $2,699. The company says it has millions of dollars in preorders and will use the money for first deliveries and more manufacturing, and it gave no shipping date. Treat that as a preorder, not a shipped product, but the price point is the useful anchor: for a short-term rental host or a small cleaning business, a $2,700 robot is a number you can weigh against a few months of part-time labor once real reliability data exists.

Behind the hardware, the race is over training data. Beijing-based GenRobot AI disclosed a $100 million Series A+ round on November 2, backed by SAIC Motor, Baidu, Galaxy General Robotics and others. It sells recording rigs that capture people's hand and body movements while they do household, manufacturing and logistics tasks, plus a platform called Matrix that cleans and labels the footage for robot makers. The investor list, which includes a carmaker and a humanoid developer, is another sign of how much of the robot supply chain runs through China.

NVIDIA, meanwhile, expanded its Cosmos tools for simulating the physical world. The open framework now supports training and fine-tuning not just NVIDIA's own models but robot models such as pi0.5, and adds new checkpoints that generate multi-camera driving video and lidar for testing self-driving systems, two-arm robot simulations and even surgical scenes. Cheaper simulation means robot makers can test more before shipping, which is ultimately what makes reliability numbers, not demo videos, available to buyers.

What This Means for Your Business

Treat AI pricing as a variable cost that is likely to rise. Anthropic's filing shows nearly $900 billion in compute commitments, and this week alone Grok tightened limits and OpenAI pooled ChatGPT and Codex into one allowance. Look at which AI subscriptions your team actually depends on, note what each plan's limits are today, and keep a second tool approved for the work that cannot stop. If a vendor offers an annual price lock, it is worth more now than it was a month ago.

Use the GDPx results to set expectations inside your company. If the best model completes about one in eight realistic expert tasks, an agent that drafts a contract, a financial model or a technical spec is giving you a starting point. Build the review step into the workflow and the schedule, and assign a named person to sign off on anything that goes to a client, a court or a bank.

Tighten your verification habits this week. The WorkOS breach shows attackers using AI to automate intrusions and then using stolen conversation history to make fraud look routine. Make it a written rule that any request to change payment details, send money or share credentials is confirmed by phone, using a number you already have, no matter how real the email looks.

If you are encouraging staff to use AI, measure outcomes, not usage. Proton's episode is a reminder that a mandate to use AI more can push people to skip the checks that protect customers. Ask each team to keep a short weekly log of what AI helped produce and what they verified by hand. That gives you real evidence of value and keeps the careful work visible.

Finally, if you are curious about robots, start a file rather than a purchase. Note price points like Lume's $2,699, ask any vendor for reliability rates and service terms in writing, and check where the hardware is made. Most of these products are a year or more from being something you can depend on, but the numbers you collect now will make the eventual decision easy.

Sources

← All posts