The 3 changes this week: A free model upgrade, an agent hub, and a 2-minute privacy audit
A smarter Claude at the same price, agents moving into HubSpot, and shared AI chats showing up on Google.
This newsletter gets drafted on a Claude scheduled task before I'm awake, so when Anthropic swaps the engine that kind of work runs on, I feel it the same morning without touching a setting. That swap happened Friday. Claude Opus 5 shipped on July 24 as the new default on Anthropic's top plans, at the same price as the model it replaced, and most of the people paying for those plans will never open the announcement.
That's one of three changes from the last seven days worth an operator's attention, and none of the three got the coverage the model-leaderboard drama usually pulls. By the end of this you'll have a two minute privacy check to run on your own AI account, a test for deciding which written-off tasks deserve a rerun, and a read on what HubSpot moving agents into the CRM means for software you're already paying for.
Here's what's in this one:
The daily-driver model got smarter at the same bill. Opus 5 replaced Opus 4.8 on Friday at identical pricing, and the right response is rerunning old failures, no new subscription involved.
HubSpot gave agents one home inside the CRM. Agent Hub and Agent Builder went to public beta for every Professional and Enterprise account on July 23.
Shared Claude chats showed up in Google results. Fixed within days, and the two minute cleanup is still worth running today.
The upgrade you already paid for
Anthropic released Claude Opus 5 on July 24 at $5 per million input tokens and $25 out, the same sticker as Opus 4.8, the model it replaces. Anthropic's own line is that it lands close to Fable 5, the company's most capable public model, at half the cost, and it's now the default on the Max plans and the strongest option on Pro. The full breakdown with benchmark tables is in tech-ish's launch coverage.
Two numbers in the launch material matter more for operators than any leaderboard placement. On AutomationBench, a test of whether a model can carry an ordinary business task from start to finish, Opus 5 passed 26 percent, against 18.1 percent for OpenAI's GPT-5.6 Sol. Zapier CEO Wade Foster says it topped Zapier's own automation leaderboard and cleared a customer-churn task no previous model had passed. Read those two together and the picture is that the best model on the market still completes about a quarter of ordinary business tasks end to end, which is why every issue in this series keeps landing on the same design: bounded jobs, human checkpoints, the Book Line and its cousins.
Here's the operator move, and it has nothing to do with switching tools. Somewhere in your head is a list of jobs you handed an AI in 2025 that came back wrong, the messy spreadsheet cleanup, the proposal draft that missed the point, the report that invented a number, and each of those failures hardened into a conclusion about what AI can't do. I call that conclusion The Stale No. Model capability moves every few months, your conclusions mostly don't, and a no you collected from a model two generations back is expired evidence you're still acting on.
When the default engine in a tool you already pay for gets a real jump, upgrade day is the day the Stale No list gets rerun. Write the list down this time:
MY STALE NO LIST
Tasks I tried with AI that failed, with the date I gave up:
1. [task] - [month/year] - [what went wrong, one line]
2. [task] - [month/year] - [what went wrong, one line]
3. [task] - [month/year] - [what went wrong, one line]
Rerun rule: when the model behind my main AI tool gets replaced,
rerun the top 3 with the same inputs I used the first time.
Keep the output only if I can verify it in under 5 minutes.
Worth flagging: every benchmark figure above is Anthropic's own run, no independent numbers exist yet, and requests that trip a safety filter quietly fall back to the older Opus 4.8, so some of what you get on any given day is still the previous model. The rerun costs you twenty minutes either way, and the worst case is confirming your no is still fresh.
HubSpot gave its agents one home
HubSpot launched Agent Hub and Agent Builder in public beta on July 23, live now for every Professional and Enterprise account. Agent Hub is one screen showing every agent running across marketing, sales, and service, live status and performance included, and Agent Builder lets you describe an automation in plain language and have it built against the deal history, contact records, and call transcripts already sitting in your CRM.
The problem it targets is real. HubSpot product chief Duncan Lennox describes what happens once a team runs more than one agent: "they become fragmented, all working from different pictures of the customer, or even worse, no picture at all." A prospecting agent emails an account the same week a service agent is handling that account's open complaint, and neither knows about the other. If you run HubSpot, the pattern I'd copy comes from the launch customer: Ignite Reading, a tutoring program operating across 25 states, pointed Agent Builder at one narrow grunt task, parsing each school district's academic calendar, and turned a 15 to 20 minute chore into seconds, north of 350 hours a year back. HubSpot picked that story for the press release, so hold it loosely, but look at its shape: bounded, verifiable, zero customer contact.
The skeptical read comes from Gartner's Kathy Ross, talking to CX Today: "They're very powerful tools, but they're not employees, they're not teammates, and they have to be managed like technology." A human rep having a bad day affects a handful of conversations, a misfiring agent reaches your whole list before anyone looks up. Which is why the first agent to turn on is one whose mistakes stay inside the building, and why the access question from the Anthropic security piece applies here word for word: what can this agent touch, and whose login is it borrowing to touch it.
The 2-minute check on your shared AI chats
Over the weekend, Reddit users found that a Google search along the lines of site:claude.ai/share surfaced long lists of publicly shared Claude conversations and Artifacts. Futurism found "a detailed medical report of a real patient, clinical trial results that included patient names," and company documents marked internal-use-only in the exposed set. Anthropic's position, via spokeswoman Amie Rotherham: "When someone shares a conversation, they are making that content publicly accessible, and like other public web content, it may be archived by third-party services." By Monday afternoon the search results were gone, per TechCrunch's own testing.
The fix on Anthropic's side landed fast. The lesson on your side is permanent, and it applies to every AI tool with a share button: a share link is a publish button with a delay on it. The chat you sent a client in March with your pricing logic in it, the thread you shared with a contractor that had customer names in it, those are public web pages, and whether they stay obscure depends on nobody ever posting the URL anywhere a crawler can see it. ChatGPT went through its own version of this last year, when a researcher scraped around 100,000 shared conversations that had been set public.
Run the cleanup now:
THE SHARE-LINK AUDIT (2 minutes)
Claude: Settings -> Privacy -> Shared Chats
Delete every link you don't have a live reason to keep.
ChatGPT: Settings -> Data Controls -> Shared Links
Same rule. If you can't name who still needs it, kill it.
Going forward, one question before sharing any chat:
"Would this page be fine showing up in a Google result
with my name on it?" If no, copy the text into a doc instead.
Where I'd start
Ranked by effort against payoff: the share-link audit is two minutes and closes a real exposure, so it goes first, today. The Stale No rerun is twenty minutes this week, and it's the one most likely to hand you a working automation you'd already paid for and walked away from. The HubSpot pilot is for HubSpot shops with an hour to spare, one bounded internal task in Agent Builder, mistakes that stay inside the building, then a look at Agent Hub's performance screen after a week to see what the thing did while you weren't watching.
I've used AI every working day since February 2023, and weeks like this one are the argument for running a system instead of a scramble: the tools underneath you upgrade themselves, sprout agents, and spring leaks whether you're watching or not. Keeping up turns out to be a set of small habits, a list you rerun, a settings page you check, a pilot you scope deliberately small. Inside the Abra AI community that's the standing conversation, operators comparing what cleared their own Stale No lists and which agent pilots earned a second week, and the skill files behind this newsletter's own automation live there too.
Recap
Opus 5 replaced Opus 4.8 at the same price on July 24, so rerun the tasks you wrote off instead of shopping for new tools. HubSpot put agent building and agent monitoring inside the CRM on July 23, and the first agent worth turning on is one whose mistakes can't reach a customer. And your shared AI chats are public web pages, so spend the two minutes deleting the ones that have no reason to exist.
Reply with the one task you wrote off as too much for AI in 2025. I want to see how many of those come back from the dead this quarter.
Andrew
P.S. Three related builds from the archive: the 4-question agent access check borrowed from Anthropic's own security team, the calendar workflow with the Book Line built in, and the three-prompt Google Business Profile check. The full archive lives at muddventures.substack.com.