Slack replies in seconds. Not minutes.
Dictate into Slack, email, LinkedIn, or any app and get polished, send-ready text. Wispr Flow strips filler and formats everything. 89% of messages sent with zero edits. Works on Mac, Windows, and iPhone.
I spent part of the other day refreshing ChatGPT for access to Astra.
A few hours later, I was using it to connect tools, test workflows, examine business problems, and build.
People call something a "ChatGPT moment" when a technology's potential becomes tangible enough to change expectations and invite much wider use.
For me, Astra is beginning to create that feeling inside ChatGPT again. The first shift changed what I believed I could ask a computer. This one is changing how much work I am willing to hand it.
The signal
1. Astra is expanding what we can delegate
OpenAI introduced GPT-6 Astra, its new general-purpose model, on September 3. The launch emphasizes computer use, research, software engineering, and professional work. It can work across tools and incorporate changes while staying with the original task. Access is rolling out in stages. OpenAI's announcement
Whenever a substantially more capable model arrives, I revisit my current workflows, my AGENTS.md files, and my business processes. That was the first thing I did with Astra. Those files contain the instructions I give my agents. Some of those instructions, and the processes around them, may reflect limitations of earlier models. I want to test which steps can be simplified, which handoffs can disappear, and which responsibilities I can now delegate. A model upgrade is a reason to reconsider how the work gets done.
A model becomes more useful when it can carry a larger piece of a project. The work includes gathering context, acting on it, checking the result, and handling your revisions.
How others are using it: Five examples show the range, from daily operations to original research.
In a company-reported evaluation, Legora used Astra to review financial statements across 41 documents. It found all four deliberately planted errors, including a £500,000 discrepancy in a revenue note, and produced a record of the checks. The expert still owned the review. Legora case study
Thomas Ricouard used Astra to create a furnished house as an editable Blender model, then turn it into an Unreal Engine walkthrough. He could revise the space and inspect it from inside. This was an iterative visualization project, not a construction-ready design. His project
Configuring a business workflow inside the software itself. At ChatPRD, Claire Vo used Astra in Codex to operate her CRM's workflow builder. It added and connected the steps to route leads, generate personalized emails, and send drafts to Slack for review. Her demonstration shows an agent turning a business instruction into a working flow inside an existing tool. Her walkthrough
Building a universe you can explore. Ricouard also used Astra in Codex to build Void Explorer, a space game with 2,048 star systems and more than 10,000 procedurally generated planets. You can fly toward a distant world, land, get out, and walk. The build combined code, terrain, rendering, 3D assets, and tests, with Ricouard directing and refining the experience. See the game and build process
Contributing to original mathematics. OpenAI published a paper crediting Astra with a proof that infinitely many pairs of prime numbers are no more than 186 apart, improving an established bound. That is a new research result for mathematicians to examine. The paper and verification materials are public; the supporting computer-checked formalization still relies on stated assumptions. Read the paper
The opportunity: Reconsider work you have kept fragmented because carrying it across several tools was too much effort. A useful test is whether AI can take one assignment from context to a reviewable result with fewer handoffs.
2. AI video is becoming something we can influence
The video breakthrough starts with a simple number: five seconds of footage generated in approximately three seconds, according to fal.
That is the performance fal reports for H3 Max, its post-trained version of MiniMax H3, introduced on August 26. The model generates synchronized video and audio. H3 Max announcement
The more recent H3 Max Turbo pushes speed and cost further. H3 Max Director supports a continuing video stream that can respond to new prompts. These are video-generation models, with early developer limits, including restricted access to longer Director sessions. Turbo, Director
When the next scene can be generated while you watch the current one, a choice or question can shape what appears next.
The video does not have to be completely decided before you press play.
How people are using it: Pieter Levels built Infinite Slop, where audience messages feed into later generated scenes. Live Classroom, an open-source project by internetphysics, turns a topic into an educational cartoon, rendering short lesson scenes near playback time. It requires the user's own fal key. Infinite Slop, Live Classroom
Filmmaker Henry Daubrez demonstrated THIS WAY, a film engine that prepares two possible futures while viewers vote on which one becomes the story. He describes ongoing problems with consistency and generation failures; it is a demonstration, not yet a publicly available product. Creator's demonstration
What I am learning from testing video: Speed alone does not create a good experience. I have been working through repeated opening shots, conversations that are too short, voices that need more character, and gaps between scenes. Story logic can be stronger than the footage carrying it.
The opportunity: As someone who teaches, I keep thinking about a lesson that changes its next demonstration when a learner says, "I still don't understand." Live Classroom demonstrates topic-to-cartoon generation. Responding to a learner's confusion would be a next step worth building and testing.
How I am using it: I have put two video experiments online.
Founder's Journey: Miami is a playable story about building a company in Miami. You watch a scene, make a business decision, and see consequences for the company, your cash, and your energy.
The launch version used fal's H3 Max Turbo for the scenes. I have continued refining the story, characters, voices, and transitions since then.
An early MiamiFounder playthrough, captured September 5.
ReelTime explores an AI-generated sitcom with viewer participation and product placement. The idea is a shared story an audience can help steer, with products appearing inside scenes and linking out to sellers.
Its first public pilot aired generated picture and audio and accepted a viewer vote. The ongoing show format is still in development; the site is online, but it is not broadcasting continuously.
A still from ReelTime's animated pilot.
Thank you to fal for the credits that helped me experiment.
My read
Now bring these developments together.
Astra can help build the software and workflows. Video models can generate the scenes that software presents.
That is how I built and iterated MiamiFounder: GPT-6 Astra in Codex for the product work, fal for the video. When repeated choices risked generating the same footage again, I asked Astra to build a shared library so players taking the same path could reuse a scene.
The experience needs more than footage. It needs rules, memory, decisions, and consequences. Connecting those pieces is where a generated clip becomes part of a product.
For more ideas, the first useful conversation can happen around a working prototype. People can experience the idea and tell you where it breaks, whether it matters, and why they would return.
The cost of testing an experience, and finding out whether anyone wants it, can fall. That may be more consequential than producing more content.
The human side
There is a tension here that connects to the first issue of this letter: access gets easier, and judgment becomes more consequential.
When starting becomes easier, it becomes tempting to start everything. More capability can fill your week with more unfinished projects.
I feel that pull. There are more things I want to try than I can give meaningful attention to.
Every product we put online still makes a promise to someone. It should work, respect their time, and offer a reason to return. Every workflow we delegate still needs someone who understands the result well enough to stand behind it.
My ambition is to use these tools to finish more of the work that deserves to exist, and become more selective about everything else.
The question
What have you ruled out because building the first usable version seemed too expensive or too complicated?
Would you make the same decision today?
To review your own setup, share your workflow notes and agent instructions with the model, then try this prompt:
Review my workflows, business processes, and agent instructions, including AGENTS.md where available, for assumptions tied to older models. Verify this model's capabilities, then rank three changes worth testing. For each, show what to change, the expected benefit, and a small test with a measurable result. Preserve business constraints and human approvals. Ask for missing context instead of guessing.
Reply and tell me. I would especially like to hear what you notice after trying MiamiFounder or ReelTime.
Until next week,
Burhan