Agents of Work
Let's Talk
October 7, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 7, 2026

Summary

Powerful AI is getting cheaper to own and easier to reach from the tools a business already uses. Mistral will hand out a trillion-parameter model by month's end, Claude now works inside Google Docs and Sheets, and Meta and Sierra published rules for how personal agents deal with businesses. The open question for owners is who controls what those agents are allowed to do.

Highlights

  • Mistral Large 4's weights arrive by month's end, giving businesses a near-frontier model they can run in-house.

  • Mistral's preview costs $1.36 per million input tokens; price an in-house option before renewing AI contracts.

  • Claude now edits Google Docs, Sheets and Slides directly; keep "ask before edits" switched on.

  • Startups under five years old can get a free year of Claude Team plus $1,000 in credits.

  • Shopify, Stripe and Walmart back a new standard for customer agents; merchants should plan for agent shoppers.

  • Anthropic's vetted Red Team tier lifts its cyber safety blocks; security teams can apply for a lighter defensive tier.

  • Half of consumer AI spending comes from the top 10% of payers, mostly people using it for work.

  • Einride now runs cab-less electric trucks daily in Ohio, watched by a remote operator.

  • Collecting robot training data fell from about $72 to $10 an hour in seven weeks at Pantheon.

Quick Takes

  • OpenAI sells a yes-or-no machine. OpenAI's Decisions API is now in public beta for all developers. Instead of writing a paragraph, it returns a probability that a statement is true, a pick from a list of options with a confidence score, or a score against a rubric. It runs on GPT-6 Luna, OpenAI says it is up to 10 times faster than the same model through its standard interface, and it costs $0.10 per million input tokens with no charge for output. Early developers noted there is no caching yet, which makes high-volume sorting less of a bargain than it looks. Read more

  • Google's search model fits on a phone. EmbeddingGemma 2 is a free, 740-million-parameter model that turns text, images, audio, video and code into a shared format for search, so a voice memo can find a video clip. It runs on a phone in about 567MB of memory and is licensed under Apache 2.0, which allows commercial use. For a small firm, that means searching its own files and recordings without sending them to a cloud service. Read more

  • Hark's agent runs errands for $20 a month. Hark, the Brett Adcock startup that raised $700 million at a $6 billion valuation in May, released Hark Pro, an agent that buys groceries, books cars, pays bills and cancels services, a year before its first devices ship. It has free, $20 and $100 monthly tiers and arrives four weeks after Meta's Muse agent launched at identical prices. Europe is not included yet because of the region's strict payment authentication rules. Read more

What's Covered in Featured News

  • Mistral Large 4: what Europe's trillion-parameter open model offers, what it costs and why its security scores matter.

  • Claude moves into Google Workspace: how the new sidebar edits Docs, Sheets and Slides, and Anthropic's free year for startups.

  • A rulebook for personal agents: the Personal Agent Protocol from Meta and Sierra, who signed on and the rival standard from Visa.

  • Anthropic opens its security models: the three tiers of vetted access and the vulnerability counts behind them.

  • Who actually pays for AI: what card-receipt data says about the heaviest consumer AI spenders.

  • Physical AI: Waymo's driverless cars in Detroit, Einride's cab-less trucks in Ohio, Avride's campus robots, Minerva's remote-controlled humanoid and Pantheon's cheaper robot training data.

Featured News

Mistral will give away a trillion-parameter model

Mistral AI, the Paris company that has become Europe's main answer to the American labs, introduced Mistral Large 4 on October 6. It is the largest model Mistral has built: one trillion parameters in total, of which 49 billion are used for any given request. That design, called mixture of experts, lets a very large model run at the cost of a much smaller one. It handles text and images natively, can answer quickly or reason step by step, and supports more than 160 languages, including every official language of the European Union. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell chips in its own European data centers.

The important part is what happens next. Large 4 is available now in a public preview through Mistral's developer console, and Mistral says the weights, the files anyone needs to run the model on their own hardware, will be released by the end of October. Preview pricing is $1.36 per million input tokens and $4.18 per million output tokens. Mistral also says its European deployment runs independently under European law, a pitch aimed squarely at companies wary of sending data to US providers.

Mistral leads with security scores rather than chat quality. The model solved 93% of the challenges in Cybench, a test built from hacking competitions, scored 82% on a test of reproducing known software flaws, and placed in the top five of the Artificial Analysis Cyber Index. "Defending software often starts with proving that a flaw is real, exactly the kind of work safety filters in closed models can block," the company wrote. On business tasks it scored 59.9% on AutomationBench, which runs routine workflows across tools such as Gmail, Google Sheets, Slack and Salesforce, and 61.7% on the DeepSWE coding test.

For a business owner, two things follow. First, a model near the top tier that you can download and run without paying per use changes the math on keeping sensitive work in-house, for anyone with the hardware or a hosting partner. Second, the same security skills that help defenders will be available to anyone once the weights are out. Expect more automated probing of small-business websites and software, and patch accordingly.

Claude moves into Google Docs, and startups get a free year

Anthropic opened Claude for Google Workspace as a public beta for every paid Claude plan. A sidebar sits beside an open Doc, Sheet or Slides file and can read, select and edit it in place. In Sheets it writes formulas, builds pivot tables and native charts, and cleans data. In Slides it builds slides from existing layouts and flags overlapping or hard-to-read text. A separate set of connectors lets users create and edit Google files from the Claude chat window.

There are two modes. The default, "ask before edits," shows each change as a suggestion card for a person to approve. "Accept all edits" lets Claude keep working until the task is done. File access follows the Google sharing permissions already in place, and on Team and Enterprise plans an owner or administrator has to switch the connectors on first. For businesses that live in Google Workspace, this removes the copy-and-paste step between the AI and the document, which is where much of the time went.

Anthropic also expanded its program for young companies. Startups founded in the past five years, or funded in the past two, can apply for a free year of Claude Team with up to five premium seats, $1,000 in credits for building on Claude, and office hours with its applied AI team. "We believe the benefits of AI will reach most people through the companies that build on top of models," Anthropic said.

A rulebook for agents that shop and book on your customers' behalf

Meta and Sierra, the customer-service AI company, published the Personal Agent Protocol, an open standard for how a consumer's AI agent deals with a business. The consumer decides what the agent may do. The business sets what agents may do on its systems, and gets a record of what they did. An agent can visit as a guest or, after the customer signs in through the familiar OAuth login process, act on the customer's account. It can reach a company through its website, its programming interfaces or its own customer-service agent, and the session carries across all three.

Genesys, Instinct, Rocket, Shopify, Stripe and Walmart have signed on. "Personal Agent Protocol helps establish how agents and merchants can work together," said Shopify's Mani Fazeli. The first draft specification is due later this month, and payments are listed as a future addition, not part of the first version.

It is not the only standard. Visa announced its Trusted Agent Protocol in June, aimed at letting agents buy from about 175 million merchants, and Shopify and Stripe back both. No European retailer, bank or payment company is named on either. For merchants, the signal is that agent traffic is being formalized by the platforms that already run your checkout. When your storefront provider offers agent settings, decide what an agent may do as a guest and what requires a signed-in customer.

Anthropic opens its strongest security models to vetted defenders

Anthropic merged its earlier programs into a single Cyber Verification Program with three tiers. Defense Access is for security teams at companies, nonprofits and government bodies, smaller security firms, open-source maintainers and individual researchers with a record of reporting flaws, with approval in a few days. Red Team Access adds authorized attack testing for organizations only and takes a few weeks. Specialized Access covers testing of safety-critical systems such as flight software, power grids and telecom networks, reviewed with the US government.

The numbers explain the urgency. Partners found 129,000 verified vulnerabilities between April and July, Anthropic's own scanning of open-source code found 5,500 more, and over 33,000 were rated critical or high severity. Partners including Booz Allen and Comcast said the models increased their rate of vulnerability finding "by months or even years." On Anthropic's own test of 50 attack scenarios, Defense Access was blocked on 46; Red Team Access was blocked on none. Verification, not the model, is now the gate that matters.

Who actually pays for AI

New data from Andreessen Horowitz, drawn from YipitData's panel of US e-receipts, shows how lopsided consumer AI spending is. The top 1% of payers average $903 a month and account for 19.5% of observed spending, up 80% in 18 months. The top 10% account for about half. The median payer spends $25 a month, and only 4.5% of eligible US consumers paid for AI in August, up from 2.1% a year earlier. The heavy spenders are mostly people buying AI to get work done: they pay for automation tools such as n8n and creative software such as Figma and HeyGen. Twenty-nine of the top 50 vendors by spending did not appear in a16z's web or mobile traffic rankings, a reminder that the tools making money are often not the ones making headlines.

Physical AI

Driverless service keeps spreading to colder, harder cities. Waymo began fully autonomous driving in Detroit on October 6, covering downtown and neighborhoods including Corktown, Midtown, Hamtramck and Indian Village. Riders for now are Waymo employees only, and the company says public service will follow "in the near future" before a gradual expansion across the metro area. Detroit's winters will make it a real test of whether robotaxis can handle snow and ice as routinely as they handle sunshine.

Freight is further along than passengers in one sense: no cab at all. Einride and EASE Logistics are running two cab-less, driverless electric trucks on a daily schedule between EASE warehouses in Marysville, Ohio, on company property and local public roads. A remote operator watches each truck and can step in. Operations began in September as part of an $8.8 million truck automation corridor project run by Ohio and Indiana transportation agencies. "This deployment reflects the maturity of our purpose-built autonomous stack, combining driverless operation with remote oversight," said Einride chief technology officer Henrik Green. Short, repetitive yard-to-yard runs are the realistic first job for these trucks, and they are the same runs many regional distributors pay drivers to do.

Delivery robots are becoming a catalog item for institutions. Sodexo, which runs food service at about 425 US colleges, agreed to offer Avride's sidewalk robots across its campuses, promising rollout "in weeks instead of months." Avride says its robots have completed more than 600,000 deliveries and already work at Ohio State, Indiana University Bloomington, the University of Arizona and Salisbury University. Fees were not disclosed.

Two smaller stories show where humanoids are heading first. Minerva Humanoids raised about $10 million in a pre-seed round led by General Catalyst for Roger, a humanoid a specialist controls remotely through a virtual-reality headset and force-feedback gloves, so the person stays outside oil and gas sites, bomb disposal or hazardous-material scenes. Its first pilot starts this month. And Pantheon, which collects training data for robot hands, says it cut the cost of a recorded hour from $71.92 in early August to $10 by late September, against $60 to $150 an hour from outside vendors, by building $650 handheld kits and scaling to 90 operators who gathered more than a million unique tasks in eight weeks. Cheaper data is what eventually makes robots cheaper to teach new jobs.

What This Means for Your Business

Revisit where your sensitive AI work runs. With Mistral Large 4's weights due this month and small models like EmbeddingGemma 2 running on a phone, keeping contracts, customer records or financial files in-house is becoming realistic for a business with a capable IT partner. Ask that partner what it would cost to host an open model for one sensitive workflow, and compare it with what you pay a cloud provider each month.

If your team works in Google Workspace and pays for Claude, pilot the sidebar on one recurring job, such as a monthly sales sheet or a client deck. Keep "ask before edits" on, and have an administrator decide which staff get the connectors before anyone switches on "accept all edits." Check sharing settings first, because Claude will be able to reach whatever each person can already open. If your company is under five years old or recently funded, apply for Anthropic's free Team year before paying for seats.

Prepare your storefront for agent customers. Ask Shopify, Stripe or whoever runs your checkout how they plan to support the Personal Agent Protocol and Visa's Trusted Agent Protocol. Decide now what a guest agent may do, such as check stock and prices, and what requires a signed-in customer, such as placing an order or changing a delivery address.

Assume attackers get the same tools as defenders. A downloadable model that scores 93% on hacking challenges will be used against small-business websites and plugins. Turn on automatic updates, review which old plugins and accounts you still need, and if you have a security provider, ask whether it has applied for Anthropic's Defense Access tier or uses similar tools.

If you move goods, watch the short runs. Cab-less trucks shuttling between warehouses and robots delivering on campuses are early, but they target the repetitive local trips that are hardest to staff. Note which of your routes are short, fixed and repeated, because those will be the first candidates when regional carriers start offering autonomous service.

Sources

← All posts