GPT-6 ASTRA: BETTER AT CARRYING WORK ACROSS YOUR BUSINESS
GPT-6 Astra is OpenAI’s newest model, released just yesterday to a limited group of organizations with a broader rollout planned over the coming days.
All of the stats say that Astra can carry work across browsers, documents, spreadsheets, presentations, code, and the other software at a remarkably faster and more consistent rate than anything before it. I change direction in the middle of projects constantly, so I also care a lot about whether a model can keep its bearings when the assignment evolves.
Access is still rolling out, it is expensive, and independent testing suggests the improvement depends heavily on the job. The early strength seems to be work it can carry through a real software environment.
If you want to go deeper, I recommend this hands-on review written by folks who had a few weeks to test the model before its release.
CLAUDE FABLE 5.1: BETTER AT STAYING WITH THE HARD STUFF
Claude Fable 5.1 is Anthropic’s new model for unusually demanding, long-running projects. It is available now to Pro, Max, Team, and Enterprise customers.
The important part for most businesses is endurance. Fable is built for work that takes hours, crosses several steps, and ends with something you can review, without needing somebody to nudge it every four minutes.
In this review, the team at Every found it fast, capable, and much easier to work with than the first Fable. They also caught it missing hard limits and inventing quotations during a source-based writing assignment.
So yes, the smartest models are getting much better at doing the work. They are also still perfectly capable of making a confident mess that takes a smart person to catch.
WE’RE BUILDING AN AI TALENT STEWARD
One of our clients (actually several, but we’re just talking about one for now) just landed on the 2026 Inc. 5000 Fastest Growing Companies list. They are (shocker) growing quickly, adding people, and finding out what most growing companies eventually learn: a hiring process people can carry around in their heads stops working when there are more people, more roles, and more decisions.
Talent acquisition was this week’s prime example.
Applications and resumes were moving through inboxes, with a very busy senior leader sorting and forwarding many of them. Job descriptions existed, but they sometimes had to do three different jobs at once: explain the role, protect the company, and somehow attract a great candidate. Important approvals lived in meetings, messages, and people’s memories.
None of that is especially unusual. Good people keep processes moving through effort, tribal knowledge, and a suspicious amount of email forwarding.
So we are helping this client design a more dependable talent-acquisition system, from an approved opening through an accepted offer. It’s not an AI system that handles all your hiring for you - actually it’s the opposite. It’s leveraging the processes you have and the technology that exists to create the “Air Traffic Control” of talent acquisition for a company that could easily grow 40% in head count in the next year.
The version we are designing has one conversational front door, which we call a Talent Steward. The goal is for a leader to ask, “Where does this opening stand?” and get an answer based on the record: the current stage, the owner, what evidence is missing, and the next action the process allows.
Behind that front door, we’re defining smaller, repeatable skills. One helps turn a loosely defined need into a clear role charter. Another checks whether the role is ready to post. Another maps applicant evidence to the approved role criteria so a person can review it consistently. Others help prepare interview guides, decision packets, and the paperwork needed for an offer.
Then we give the system context: the company’s goals, values, leadership structure, hiring process, role requirements, and examples of what good work looks like.
We also give it boundaries. The AI can draft, organize, map evidence to an approved rubric for human review, and point out a missing approval. It cannot authorize headcount, publish a job, reject or advance a candidate, choose the hire, approve an offer, or send a message pretending to be a person.
The AI gives the process structure, clarity, accountability... but people still make the call.
We’re literally building the system as I’m typing this. Multiple AI tools are being used to build it, check our logic, and check each other’s work. Our development team will weigh in soon before a version 1 is trailed. All of this went from “what if” to “why not” in less than a week. That’s the moral of the story.