July 2026 put agents inside the tools teams already pay for. OpenAI launched ChatGPT Work and began charging credits for workspace agents, Claude Cowork reached web and mobile, Anthropic released Claude Opus 5, Zapier folded Agents into Zaps, and MCP went stateless. CoStar barred AI agents from its logins, and the EU delayed the AI Act's high-risk rules.
What should an operator do this month?
Three things that still apply as of late September, in order of how many teams they affect.
- If you are on Google Workspace Business Standard or Business Plus, check the Meet "Take notes for me" setting in the Admin console. It turns on for meetings with three or more people unless you change it.
- If you run ChatGPT workspace agents, set spend alerts and look at a week of credit use before rolling agents out to more staff.
- If you license CoStar, tell the team in writing that CoStar logins never go into an AI agent, browser agent or shared automation.
CoStar's new terms bar AI agents from its logins
What happened (July 6, 2026). CoStar's license terms dated July 6, 2026, which govern agreements signed from that date, say passcodes may not be shared with any third-party AI service, including AI agents. If that happens, CoStar reserves the right to raise license fees to cover the additional users it detects. Its public Terms of Use carry the same passcode rule. Source: CoStar license terms CoStar Terms of Use.
Why it matters. The shortcut many brokerages were trying, handing a browser agent a CoStar login to pull comps, now has a stated price: a bigger bill, on top of the existing limits on scraping and on feeding CoStar data to AI.
What to do. Keep CoStar work inside CoStar and automate the steps around it with your own files and public records. Can Claude connect to CoStar? walks through what is and is not allowed.
ChatGPT workspace agents started costing credits
What happened (July 6, 2026). The free preview of workspace agents ended for ChatGPT Business, Enterprise and Edu. Agent runs inside ChatGPT now draw on workspace credits, metered by input, cached input and output tokens, on top of the per-seat fee. Runs triggered outside ChatGPT, such as an agent answering in Slack, stayed in free preview. Source: OpenAI rate card, reported by Tech Times.
Why it matters. An agent that ran free all spring now has a running cost, and there is no fixed price per run. A chatty agent that writes long outputs costs far more than one that answers briefly.
What to do. Turn on spend alerts in workspace settings, pick the cheaper model for routine agents, and review a week of usage before adding users.
Claude Cowork reached web and mobile, with tasks that keep running
What happened (July 7, 2026). Anthropic brought Claude Cowork to the web and to the iOS and Android apps in beta, starting with Max. Sessions run on Anthropic's servers, so work and scheduled tasks continue with no device online, and Claude asks for decisions on your phone. The Microsoft 365 connector gained write tools for email and calendar the same day. Source: Claude blog Claude release notes.
Why it matters. An owner can hand off a recurring task, such as a 6 a.m. client prep pack, and check it from a phone. Write access to Outlook means the agent can send and schedule, not only read.
What to do. Decide who may give Claude write access to email and calendars before anyone connects it. See what Claude Cowork does.
Since then. On September 16, 2026 Anthropic began merging Cowork and chat into one Claude app, rolling out to Pro and Max first with Team and Free to follow. Source: Claude blog.
OpenAI launched ChatGPT Work, an agent that delivers finished files
What happened (July 9, 2026). ChatGPT Work, powered by GPT-5.6, takes an outcome, gathers context through plugins for tools such as Slack, Microsoft Teams, Google Drive and SharePoint, and returns documents, spreadsheets, decks, reports and websites. It includes a plan you approve before work starts, check-ins and action approvals. It rolled out first to Pro, Enterprise and Edu, with Plus and Business following, alongside a new desktop app. Source: OpenAI, corroborated by Bloomberg via BNN.
Why it matters. ChatGPT and Claude now both sell the same idea: give the agent a job, not a prompt. The approval controls decide how much it does without asking.
What to do. Test it on one recurring deliverable your team already produces, with plan approval on, and compare the result with the human version. Claude vs ChatGPT for business compares the two.
Since then. On September 22, 2026 OpenAI released GPT-6 Sol and GPT-6 Luna, available in ChatGPT Work for Plus, Pro, Business, Enterprise and Edu. Source: OpenAI.
Claude added admin controls for keys, users and HIPAA
What happened (July 8 and 14, 2026). API keys and Admin API keys can now be given an expiry date in the Claude Console. Enterprise organizations got Admin API endpoints, in beta, to list, change and remove members and manage invites and groups. HIPAA configuration became self-serve for eligible organizations. Source: Claude Platform release notes Claude release notes.
Why it matters. A key that never expires is the most common loose end when a contractor leaves. Health-adjacent practices no longer wait on a sales call to set up HIPAA.
What to do. Put an expiry on every new API key and rotate the old ones. AI governance lite has a short access checklist.
Zapier folded Agents into the Zap editor
What happened (July 15, 2026). AI by Zapier with looping tool calls became generally available, and Zapier began moving its standalone Agents product into it. Each agent is converted automatically into a Zap with a trigger and one AI by Zapier step, billed with the same task-based pricing as the rest of Zapier, and runs over 75 tasks pause for review. Knowledge sources did not carry over at launch. Source: Zapier Help Center.
Why it matters. There is no separate Agents subscription to budget for, but agent work now spends the same task allowance as your Zaps.
What to do. Open each converted Zap, check the trigger fired as the old agent did, and rebuild any agent that relied on a knowledge source. Zapier vs AI agents covers when a Zap is the right tool.
Since then. Enterprise trial accounts that had not migrated converted to free Agents accounts on August 15, 2026, as the same help article set out.
Google Meet's note-taker defaults on for Business plans
What happened (July 16, 2026). Google added an admin setting that turns on "Take notes for me" automatically for meetings with three or more participants. It is on by default for Business Standard and Business Plus and off by default for Enterprise Standard, Enterprise Plus and Frontline Plus. Source: Google Workspace Updates.
Why it matters. Small businesses are the ones on Business plans, so they get automatic AI notes on client calls unless someone changes a setting. That matters for confidential calls and for consent.
What to do. Decide whether client meetings should be summarized by default, set it in the Admin console, and tell staff how to turn it off per meeting.
Since then. On September 22, 2026 Google said the settings begin affecting users on September 29, 2026. Source: Google Workspace Updates.
Anthropic released Claude Opus 5
What happened (July 24, 2026). Opus 5 became the default model on Claude Max and the strongest selectable model on Claude Pro, and shipped on the API, Amazon Bedrock, Google Cloud and Microsoft Foundry at $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. Effort settings let teams trade capability for cost. Source: Anthropic.
Why it matters. Teams on Pro and Max got a stronger model at no extra cost. API users got a better model at the old price, with effort as a new lever on the bill.
What to do. For API work, test your prompts at a lower effort setting first; many routine tasks do not need the maximum.
Since then. On September 22, 2026 Anthropic released Claude Opus 5.5, which it says costs about 40 percent less than Opus 5 on typical workloads. Source: Anthropic.
The EU delayed the AI Act's high-risk rules
What happened (July 24 and 27, 2026). The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on July 24 and entered into force on July 27. It moved the main obligations for stand-alone high-risk systems, such as AI used in hiring, from August 2, 2026 to December 2, 2027, and for AI built into regulated products to August 2, 2028. Source: EUR-Lex, summarized by Hunton.
Why it matters. A US business with EU customers or staff that uses AI to screen candidates has an extra sixteen months. The rules were deferred, not dropped.
What to do. If you sell into the EU, keep a list of where AI touches hiring, credit or other decisions about people, and check whether the transparency rules for chatbots and AI-generated content cover what you deploy or publish there.
Since then. The transparency obligations began to apply on August 2, 2026, as scheduled; marking duties for AI systems already on the market before then apply from December 2, 2026. Source: Hunton.
The MCP specification went stateless
What happened (July 28, 2026). The 2026-07-28 revision of the Model Context Protocol removes the initialize handshake and session IDs, so each request stands on its own, and moves client registration from dynamic client registration to Client ID Metadata Documents. Roots, Sampling and Logging are deprecated with a twelve-month minimum window. The TypeScript, Python, Go and C# SDKs support it. Source: MCP blog.
Why it matters. Stateless servers are cheaper to host behind ordinary load balancers. Older servers and clients will need upgrades over the next year. What is an MCP server? covers the basics.
What to do. Ask whoever runs your MCP servers which revision they support and when they will move. Our September 2026 edition covers the revision in more detail.
An item makes the digest only if it changes a decision a small or mid-size team would make, and only with a dated source inside the month. Model benchmarks, leaks and funding news without a product consequence are left out.
Common questions
What were the biggest AI changes for business owners in July 2026?
OpenAI launched ChatGPT Work on July 9 and started charging credits for workspace agents on July 6. Anthropic brought Claude Cowork to web and mobile on July 7 and released Claude Opus 5 on July 24. Zapier folded its Agents product into Zaps on July 15, and the MCP specification went stateless on July 28.
Can I connect an AI agent to my CoStar login?
No. CoStar's license terms dated July 6, 2026 say passcodes may not be shared with third-party AI services, including AI agents, and let CoStar raise license fees for extra users it detects. Keep CoStar work inside CoStar and automate the steps around it with your own files and public records.
What did the EU AI Act change in July 2026?
The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published on July 24 and took effect on July 27, 2026. It moved the main high-risk obligations from August 2, 2026 to December 2, 2027, and to August 2, 2028 for AI built into regulated products. The transparency obligations still began on August 2, 2026.
What happened to Zapier Agents?
On July 15, 2026 Zapier moved its standalone Agents product into the AI by Zapier step inside the Zap editor. Existing agents were converted into Zaps with a trigger and one AI by Zapier step, billed with the same task-based pricing as the rest of Zapier. Enterprise trial accounts had until August 15, 2026 to migrate.
Is this page current?
It was written in September 2026 looking back on July. Where something changed after July, such as Cowork merging into the main Claude app or Opus 5.5 replacing Opus 5 as Anthropic's newest Opus model, the item says so in a Since then note with the date and a source.