Four Labs, Six Days

There are weeks where nothing happens and weeks where everything does, and we just had one of the second kind.

Anthropic, OpenAI, Meta, and Google all shipped inside about a week. Buyers I talk to are past excited and into exhausted, which is the correct response, because you cannot evaluate four frontier releases and run a business at the same time.

So here is the version that respects your Saturday. What happened, what it actually means if you are running something small, and whether it deserves any of your attention this month. Most of it does not.

GPT-6 Astra landed on September 3 and it is not a chatbot upgrade. It is built to run long tasks across browsers, spreadsheets, and desktop applications. Operating software rather than answering about it.

The specs are serious. Roughly a million tokens of context, 128,000 tokens of output, and pricing at ten dollars per million in and fifty per million out.

That price is the story. Astra runs about thirteen times more expensive per token than Gemini 3.8 Flash. So the question is not whether Astra is better, it plainly is at the thing it was built for. The question is whether the thing it was built for is a thing you need.

What it means for you: If you have a workflow that currently requires a human to sit in front of four applications for forty minutes moving data between them, this is the first model I would genuinely try on it. If you do not have that workflow, Astra is not for you yet and using it for ordinary work is an expensive way to get an ordinary answer.

Do this: Write down the single most tedious multi application task in your week. Not to automate yet. Just to have a candidate ready when you decide to test.

Fable 5.1 went GA on September 1 at the same ten and fifty pricing as the version before it. Two things changed underneath.

Cache reads dropped to twenty five cents per million. If you run the same large context repeatedly, which describes almost every production workflow with a knowledge base attached, that is a meaningful reduction in the boring part of your bill.

And the self reported benchmark jump on Terminal-Bench-Science went from 24.7 to 52.6. Roughly doubled. Treat vendor reported benchmarks with the skepticism they deserve, but a doubling on a technical benchmark is not a rounding artifact.

There is also a trusted access twin called Mythos 5.1 that most people will never touch, which is worth knowing exists mainly so you understand why capability tiers are starting to fork by access level rather than just by price.

What it means for you: The cache pricing is the practical item. If you have an assistant that gets the same fifty page context every single run, your cost profile just changed.

Do this: Find your highest volume workflow that resends a large fixed context. Check whether your setup is actually using caching. A surprising number are not, because it was off by default when they built it and nobody went back.

Released September 2. Twenty percent fewer tool calls and twenty five percent fewer tokens than the previous version, for the same work, with a million token context window.

I want to flag this one because it is the release everybody underrated.

Every other lab this week made things more capable. Meta made the same capability cheaper to execute. Fewer tool calls also means fewer round trips, which means faster wall clock time and fewer places for an agent run to fall over. Efficiency gains are less fun to announce and they are what actually shows up in your operating costs.

What it means for you: If you run agents that chain many tool calls, the failure rate of a chain compounds with its length. Cutting calls by a fifth is a reliability improvement disguised as a cost improvement.

Do this: Nothing urgent. But when you next evaluate models for an agent workflow, count tool calls per completed task alongside cost per task. Most people only look at the second one.

Z.ai shipped the first natively multimodal GLM-5, at 320 billion parameters with 18 billion active, a million token context, and MIT licensed weights. Self reported DeepSWE went from 46.2 on the previous version to 63.4.

MIT licensed is the part that matters. You can run it yourself, modify it, and deploy it commercially without asking anyone.

What it means for you: For most readers, nothing immediately. Self hosting only makes sense if you already have infrastructure competence in house, and if you do not, learning it costs more than the API bill you are trying to escape.

But there is a category where it changes the math completely. High volume, low complexity, privacy sensitive work. Document classification on client files. Internal search across material you cannot send to a third party. If that describes something you do, the option now exists at a quality level that would have been frontier a year ago.

Do this: If you have ever declined to use AI on a workflow because the data could not leave your control, this is the week to revisit that decision.

The two announced an expanded partnership putting Claude inside Salesforce as a default reasoning model across Agentforce, Slack, and developer tooling, with something like thirty seven prebuilt sales skills and pilot access ahead of an open beta this month.

What it means for you: If you are on Salesforce, this is the most consequential item on this list for your actual day, and it is worth applying for the pilot rather than waiting.

If you are not on Salesforce, the signal still matters. The pattern of the frontier lab embedding directly into the system of record, rather than sitting beside it as a separate chat window, is where this is all heading. Expect the same shape from every major platform you use within a year.

Do this: Salesforce users, apply for the pilot and start a data mapping exercise now so your permissions and business rules are ready when the beta opens. Everybody else, note which of your platforms has not announced something like this yet, because that is a switching risk you should be tracking.

Between August 24 and September 1, Okta, Auth0, and Descope each shipped implementations of Cross App Access, a pattern for how agents get delegated permissions across applications.

Three competitors converging inside eight days is not coincidence. That is a standard being ratified in public.

What it means for you: Agents are getting their own identities, separate from the humans who deploy them. Which means the era of handing an agent your own API key is closing, and good riddance, because an audit log that says you did something an agent did is worse than no audit log.

Do this: Check whether your identity provider has shipped support. If it has, plan the migration. If not, at minimum give every agent its own service account with scoped permissions today, and never share one across two workflows.

The State of AI in 2026 survey found that thirty two percent of organizations have skipped buying at least one software product or feature because they could build it internally with agentic coding tools.

Almost a third. That is a structural change in how software gets bought.

The same survey found large enterprises scaling agents in one or more functions went from twenty seven percent to forty percent, while smaller firms stayed flat at twenty two percent. The gap is widening, not closing.

What it means for you: Two things, pointing opposite directions. If you sell software, a third of your pipeline now has a build option they did not have eighteen months ago. If you buy software, you have leverage you are probably not using.

The flat number for smaller firms is the one I keep thinking about. The tools got cheaper and more accessible and small companies did not move. That is not a capability problem. It is an attention problem.

Do this: Look at your software spend. Find the one line item that is a simple tool doing a simple thing at a price that annoys you. That is your candidate.

A report out this week counted roughly 17,800 public AI add ons across something like 6.7 million installations.

Sit with that number for a second. Most of those are browser extensions. Most browser extensions request permission to read and change data on every site you visit. And most people installed theirs during a ten minute window when they were curious about a tool somebody mentioned, and then never thought about it again.

Your browser is where your email lives, your bank lives, your ad accounts live, and your CRM lives. An extension with read access to all sites is functionally an employee with a key to the building who you never interviewed.

The risk is not usually that the extension was malicious when you installed it. It is that extensions get sold. A developer with a hundred thousand installs and no revenue gets an offer, the extension changes hands, and the update that ships next month does something the original never did. It has happened repeatedly and it will keep happening because the incentives are perfect for it.

What it means for you: You almost certainly have three or four of these you forgot about, and at least one of them can read every page you load.

Do this: Open your extensions page right now. It takes four minutes. Remove anything you have not deliberately used in the last month. For the ones that stay, look at what permissions they have, and if something that summarizes text has permission to read data on all sites, ask why. This is the highest security return per minute available to you this week and it does not require you to understand anything technical.

OpenAI's chief scientist published an essay calling for a slowdown in AI research. Jakub Pachocki joins a growing list of senior people saying this out loud. No practical implication for your Monday. Considerable implication for how you plan two years out.

Anthropic walked away from acquiring Decart at around six billion after due diligence. Deals dying in diligence is normal and usually private. This one being visible is mostly interesting as a signal that acquirers are getting more careful at the top of the market.

Alibaba Cloud, Cambricon, and Ant Group joined the PyTorch Foundation. Infrastructure governance, invisible to you, matters enormously to what your tools cost in three years.

If you did one thing from this entire list, make it the cache check on your highest volume workflow. It is fifteen minutes and it either finds you money or confirms you are already fine.

If you did two, add the service account cleanup. Separate identity per agent. It is tedious and it is the kind of thing that is trivial now and miserable in a year when you have thirty of them.

Everything else on this list is a watch item. Four frontier releases in a week is genuinely a lot, and the correct response to a lot is not to evaluate all of it. It is to notice which one touches something you already do, test that one, and let the rest keep happening without you.

The labs will ship again in three weeks. They always do. Your business does not need to move at their pace and the operators I know who are actually winning are running about a quarter behind the news and roughly three times more profitable than the people who chase every release.

FROM THE AI NEWSROOM

For operators who would rather build it with someone than build it alone. The full stack, installed alongside you, with the decisions made in the room instead of guessed at. Ninety seven dollars.

Join the Accelerator for $97

.....

Jordan

The AI Newsroom | Practical AI for people with a business to run.