What we would do: make your business completable by a machine. A form that submits, hours and prices as text rather than pictures, and a Google Business Profile that matches your site to the letter.

Google, the referee. On September 24 at 9:15 AM Pacific, Google released the September 2026 spam update, global and in all languages, with a rollout that may take up to two weeks. It is the fourth spam update this year, and Google’s own spam policies list exactly what it enforces, from cloaking to scaled content abuse.

What we would do: nothing reactive for two weeks. One look at Search Console for a manual action, one honest read of the scaled content policy against anything you generated in bulk, then keep publishing pages a customer would read.

Claude, the colleague. Anthropic’s Claude Tag can now use your own connectors for a request you make in a Slack channel: your calendar, your drive, your assigned CRM accounts. Nobody else can use them, every action logs under your account, and you decide whether answers post or wait for review. Team plans first, Enterprise next.

What we would do: one connector, in the channel where the question usually gets asked, with review on for the first week.

Plus the Claude Marketplace and what ties the three together, in the full edition: https://miningwells.com/ai-news/meta-muse-fee-google-spam-update-claude-tag-connectors

What we would do: move one workflow you were rationing on cost, a client report or a landing page draft, up to the premium model this week and see what changes.

ChatGPT, half price. Ninety minutes later OpenAI introduced GPT-6 Sol and Luna: API prices 50% below GPT-5.6 promotional pricing, Sol at $2 in and $10 out per million, Luna at $0.10 and $0.50.

What we would do: treat your AI subscription as a variable cost you renegotiate every quarter, not a fixed one.

Meta, the concierge. Reuters reported that Meta tested human contractors quietly placing some of Muse’s phone calls, then rolled it back after an internal post called the missing disclosures “a miss.”

What we would do: if any part of your business puts AI in the room, write the one sentence that says when a human is actually doing the work, before a customer has to ask.

Plus SpaceXAI’s Grok Bot absorbing a 175% support ticket spike without a hire, in the full edition: https://miningwells.com/ai-news/claude-chatgpt-price-war-muse-human-backup

What we would do: point an AI at your own unread pile (reviews, tickets, call transcripts), then have a person check ten of its findings against the originals.

ChatGPT, the ad network. ChatGPT Ads rolled out to seven more Asian markets and now run in more than 60 countries, with self-serve Ads Manager for eligible businesses. OpenAI says ads passed a $1 billion annualized run rate in under 200 days. Measurement tools are still coming.

What we would do: a small test budget while the auction is young, with ads written for the question a buyer would type.

Muse, the shopper. Meta gave Muse real-time voice, Mac computer use and connectors for Walmart, Best Buy, Sephora, Ulta, Wayfair, Shop Pay and PayPal. The most important shopper this quarter has no thumbs.

What we would do: fix the specs and answers on your ten best product pages, because the agent reads what you wrote down.

Plus Amazon putting Seller Assistant inside Claude, in the full edition: https://miningwells.com/ai-news/claude-enzyme-chatgpt-ads-asia-muse-amazon-seller-assistant

We read all three so you do not have to, with one move per story you can make this week.

Read the full edition:

miningwells.com/ai-news/grok-4-7-muse-chatgpt-amazon-block

We read all three so you do not have to, with one move per story you can make this week.

Read the full edition: miningwells.com/ai-news/chatgpt-ad-cookie-anthropic-new-model-grok-bot

Hello,

I want to tell you about a week I spent with Meta's new AI agent, because the lesson in it is not really about Meta. It is about how we decide what to depend on, and I think most of us are making that decision backward.

I wrote the long version of this on our site, with the leaderboard screenshots and the full benchmark tables: the complete Meta Muse review.

First, what happened

Meta shipped Muse on September 8. A personal AI agent with its own browser, built to do things on your behalf rather than talk to you about them.

I gave it the admin that eats my week. Signing into accounts. Editing profiles. Setting up recurring posts. Real work, not demos.

It was extraordinary. Faster than anything I have used, running several browser sessions in parallel without losing the thread on any of them, going much further on its own before it needed me. There was a moment early on where I sat and watched it work and thought, this changes my Tuesdays.

Then it started stopping. Tasks would begin, get most of the way, and just halt. Not fail, halt. I would come back to a session sitting idle and have to nudge it along. Then nudge it again.

Then, a few hours in, every browser it had went down at once and stayed down until that night. I tried from different machines. Same dead panel every time.

The error reporting was accurate, which is rarer than it sounds

When I asked what was happening, the explanation it returned turned out to be accurate.

It said its browser backend had been erroring since that morning, that this was infrastructure rather than my account, that chat and files were fine while browser automation was down. Asked whether it was affecting everyone, the reply was that no outage reports could be found and that it could not determine whether the problem was one session or something wider.

When I pushed, the explanation it returned got specific in a way I found oddly reassuring.

Muse does not browse from its own computer. It hands the job to Meta's system, which starts a real Chrome browser on one of their servers and runs the task there. What had broken was that handoff. Its requests were returning "failed before dispatch," which means they never reached a browser at all. The analogy it produced: calling a taxi dispatcher who never picks up. There is no car to inspect.

Failed before dispatch is not a browser crashing. It is requests never getting a browser assigned in the first place, which is what a system under more load than it planned for looks like.

It did not hallucinate an explanation. An LLM asked to account for its own failure will usually generate something plausible and wrong, because generating plausible text is what it does. Getting an accurate diagnosis instead means the system is surfacing real state into the model's context rather than leaving it to guess, and that is an engineering decision worth noticing.

There is a second thing worth telling you, because almost nobody is writing about it. Muse never sees your passwords. They sit in a separate credentials store, and a second system called Sentinel injects them at the network boundary after an action is approved. The model can propose an action. It cannot authorize one.

Which means if it is ever prompt-injected into attempting something it should not, there is nothing for it to leak. It never held the secret.

Early on it surfaced a notice that a stored sign-in had expired and a fresh one was needed, rather than failing quietly or prompting me to paste a password into the chat. In the moment that felt like friction. It is the architecture declining to route around an expired credential, and it is the best security design I have seen in this category.

And then it got fixed

Muse has a Report an issue form built into the app, so I used it. My entire report: "My browser is down, please help!"

A few hours later, that night, it was back. I asked whether it was working. Muse retried with a fresh browser, loaded Google, and replied: "Yeah, it's working. Google loaded clean and the browser's ready to go."

Whether my report or Meta's own monitoring got there first, I cannot say. But a problem reported inside the product was fixed within hours, and the product confirmed the fix itself. That counts, and I want it on the record as clearly as the outage.

Now the part that actually matters to you

Here is the thing I keep chewing on.

Meta solved the hard problem. Getting software to plan a multi-step task, drive a real browser, run several jobs at once, recognize a form, and halt for permission at the right points is genuinely difficult engineering, and they have it working at consumer scale in week one.

What broke was the easy part. Keeping servers up. And that got fixed the same day.

That is a much better position to be in than the reverse, and it is why I am not writing this off. But it exposed something about how I had been thinking, and maybe how you have been too.

When we evaluate a new tool, we watch the demo. The demo is the good day. Every tool is impressive on its good day, because that is what a demo is for.

The question that actually determines whether a tool is worth building a business on is: what happens on the bad day? Not whether it fails, everything fails. Whether it fails loudly, whether you find out immediately, and whether the work you handed over is recoverable.

Muse reported its failure accurately, which is worth a great deal. But it failed in the middle of work I had already delegated and stopped thinking about, which is the expensive kind of failure. The cost is not the outage. The cost is the hours you spent believing something was handled.

What I am actually doing about it

Using it, with a hand on the wheel.

My verdict: it is great. When it is running, Muse is the fastest thing on my desk, and I would use it for real client work, including outreach and social media management. The parallel execution is a real edge.

Use it at your own risk, though. In my experience it needs a lot of oversight: check in on long tasks, nudge the ones that stall, and review its work before it reaches a client.

For what it is worth, I did not leave this on impressions. I went and looked at two independent scoreboards directly, Artificial Analysis and Arena, both on September 20.

Arena keeps a separate leaderboard just for agents. Muse is not in the top ten of it. Claude Fable 5.1 leads at 13.71 percent, and seven of the ten slots are Claude models. That is a blind human preference test, not a vendor's own scorecard, and it matched my week almost exactly.

Muse is not a weak model. It sits fourth on Arena's text board and top ten on web development and vision. On Artificial Analysis it scores 48 on intelligence against 51 for Claude Opus 5 and 53 for Claude Fable 5.1 and GPT-6 Astra. Close enough.

Then you see the other two columns, and the whole product makes sense. Muse runs at 224 to 250 tokens per second where the Claude and GPT frontier models run 54 to 70. It costs $1.60 per task where Claude Opus 5 costs $5.86 and Claude Fable 5.1 costs $7.63.

Five points of intelligence, traded for four times the speed at a quarter of the price.

That is not a company failing to build a frontier model. That is a company deciding most people would rather have fast and cheap and safe than the last five points of clever, and honestly, for a lot of work, they are right.

One more thing, because the timing is too neat to ignore

Muse launched on September 8. By September 18 it was the number one app in the US App Store, above ChatGPT, Gemini, Claude and Instagram, with more than 730,000 downloads in ten days.

My browsers went down two days after that.

Think about what Muse hands every single user: not an API call, but a dedicated virtual machine running a live browser that keeps going after you close the app. That is a whole computer per person. A chat product that goes viral needs more inference capacity. An agent product that goes viral needs more actual machines.

I cannot prove that is what happened, and Meta has not said it. But it fits, and if it is right then what I hit was not bad engineering. It was a company being caught out by its own success at something genuinely hard.

It came back the same day, which fits that reading. Capacity is something you can add in hours. A broken design is not.

Which is the summary of the whole week. The idea is there, and when Muse is running, so is the execution.

The test I would run

If you are weighing Muse, or honestly any agent, do not run the exciting test. Run the boring one.

Give it a multi-step task. Walk away for an hour. Look at the state you come back to. Do that on three separate days.

Those three hours will tell you more than any feature list, any review, or any newsletter, including this one.

I am running it again in a month. Based on what Muse does when it is actually running, I expect to write you an even better letter then.

Until next time,

Jessica

The full review, with every screenshot and the benchmark detail, is here: Meta Muse review on miningwells.com.

If you have been using Muse and your week went differently, hit reply. I read everything, and a second data point is worth more to me than another feature announcement.

Sources: Meta newsroom · ai.meta.com/muse · Meta Help Center · Meta AI Research · Artificial Analysis LLM Leaderboard · Arena Leaderboard · TechCrunch, Muse reaches No. 2 in the US · Muse hits No. 1 with 730,000 US downloads

Benchmark figures read live from Artificial Analysis and Arena on September 20, 2026. Both leaderboards update frequently.

Claude now leads a quarter of the work that builds the next Claude

Anthropic published three measurements of how much AI is doing inside its own lab. As of August, Claude leads 26 percent of its AI research and development work, meaning it can take most of a task from a prompt to done while a person supervises. In February that share was under 1 percent. The method is worth as much as the number: Anthropic sampled 20 percent of its research staff every week in July, split the work into roughly 15,000 tasks, and had Claude help rate them, which it says out loud.

What we would do: stop treating “we tried AI in March and it was fine” as a settled verdict. Rerun your best prompts and workflows every quarter and keep whatever got better.

ChatGPT moved into Microsoft Word, including the free plan

On September 17, OpenAI shipped ChatGPT for Word: a sidebar that drafts from notes, summarizes, revises and cleans up formatting, on every plan including Free. Quieter item on the same page, and the more expensive one if it applies to you: OpenAI plans to retire custom GPTs, with a migration path to plugins.

What we would do: assume every competitor’s proposal, press release and blog post now comes out of that same sidebar. Average writing just became free, and free plus average is a miserable place to compete. Put your customer’s exact words, your real numbers and an actual opinion into yours.

A dime an hour to turn every sales call into copy

xAI released Grok Voice Transcribe 2.0 at ten cents per hour of recorded audio and twenty cents for live streams, which it says is twice as accurate as version 1.0 at the same price. It labels speakers, takes up to 100 custom terms per request and can drop every um. Atlassian says Loom uses it.

What we would do: transcribe your recorded sales calls, pull the exact phrases customers use for their problem, and put those phrases in your headlines, ads and FAQ answers. People trust their own words, and so do the AI answers that quote pages sounding like the question.

The pattern underneath all three

A good model is no longer an advantage, because your competitor has the same sidebar on the free plan. What separates one business from another is what you feed it: first-party data, recorded customer language, proof you can show and a point of view worth quoting. Best inputs win, not most subscriptions.

Read the full edition, with a source link on every claim. Free, no login, no email required.

Two years earlier, in Summer 2024, it was 25 of 248. About one in ten.

We built this from YC’s own public listing, batch by batch, and it is one of the two sharpest shifts in what YC funds anywhere in the last five years of data. The other one is below.

Two years of the same number, then a jump

What makes the chart interesting is its shape. It is not a steady climb.

From Summer 2024 through Fall 2025, Industrials held between 8 and 11 percent of every batch, five batches in a row. Winter 2026 moved it to 14 percent. Spring 2026 held there. Then Summer 2026 landed at 24 percent, and the first 51 companies listed for Fall 2026 are running at the same rate, 12 of them industrial.

That is what a change in direction looks like, as opposed to a trend line. Something moved between the spring and the summer.

Fintech went the other way

In Winter 2022, 89 of YC’s 398 companies were fintech. That was 22 percent of the batch, second only to B2B.

By Summer 2023 it was 7 percent. In Spring 2026 it was 11 percent, 21 of 193, and in Summer 2026 fintech no longer made the top three at all. Healthcare took the third spot.

YC did not stop funding hard problems. It moved to a different kind of hard problem.

YC is telling founders what it wants

You do not have to guess at why. Y Combinator publishes a list of the companies it wants to fund, its Requests for Startups, and the current list reads like a description of this chart.

It includes “The Future of American Defense,” which asks for things like low-cost interceptors, next-generation sensors, drones and advanced manufacturing. It includes “New Operating Systems for the Physical World,” about robots actually deployed in the field for construction, maintenance and fleet work. There is “Compute at Sea,” about offshore data centers, and “Data for the Real World,” which includes robots that collect data in places that are hard to reach.

When the most closely watched accelerator in startups starts asking for companies that build physical things, founders listen. The Summer 2026 batch is what listening looks like.

Why this matters past the batch

Here is the part that is easy to miss, and it is the part I think about most, because it is our job.

An industrial company does not get sold the way a software company does. The buyer is a plant manager, a fleet operator, a procurement team or a government program office. The sales cycle runs months or years instead of days. Nobody signs up with a credit card. A deal needs a committee to agree, and each person on it needs a different reason to say yes.

A lot of what the startup world knows about growth was learned selling software to other software companies: a free trial, a fast demo, a self-serve checkout. Very little of it survives contact with a buyer who has to justify a purchase to an engineer, a finance lead and a safety officer at the same time.

So if nearly a quarter of a YC batch is now selling to industrial buyers, a lot of founders are about to learn a different playbook: proof over polish, case studies over clever copy, and a message that works for every person in the room, not just the one who took the first call.

Go deeper

Y Combinator publishes what it wants funded next, and we read all 13 ideas on its current Requests for Startups so you do not have to, with who would actually buy each one.

If you are wondering how hard it is to get into a batch at all, here is what YC has published about acceptance, with batch sizes through 2026 and the standard deal. And if you are an industrial founder starting to raise, we keep a free directory of venture capital firms, organized by region and stage.

Nothing to sign up for. It is there to use.

Jessica Wells is co-founder of Mining Wells Marketing Agency. She is a four-time founder with more than a decade in marketing and growth, and a former consultant at Booz Allen Hamilton.

Source: Y Combinator’s public company directory, snapshot of September 13, 2026, using the industry labels YC assigns. YC’s Requests for Startups quoted from ycombinator.com/rfs.

If this was useful, subscribe. Every issue is one finding from data we compiled ourselves, in one short read.