GPT-6 Astra is the smartest and strangest model yet — coders are retreating to older models while everyone else discovers it can do things that were never part of their job. NLW turns that confounding release into a playbook: a dozen thought starters for opportunity AI, from marketing people can play to proposals clients can reshape to packaging your own judgment as a product.
The new models' value is in things you've never considered — which means you need thought starters, not benchmarks.
Broad model capability has reached a level where new releases like GPT-6 Astra often won't make your existing AI work noticeably better — sometimes they'll even feel like a regression. Their real value is unlocking capabilities you've never used: games, video pipelines, 3D, interactive experiences. But by definition, you haven't considered what you've never considered. Nobody walks around with a complete inventory of things they might make. So this episode is a set of thought starters for opportunity AI — the mindset of asking what you might do now that you never would have before.
GPT-6 Astra is in some ways far beyond anything before it, yet in familiar use cases the improvements aren't noticeable and can even feel like a step back. That's the marker of where AI development has arrived: new models' value increasingly isn't doing your existing work better, it's unlocking capabilities you've never considered.
A week in, power users describe Astra as the smartest and dumbest model they've worked with — mood swings, absurd shortcuts to something technically working. OpenCode's team partially reverted to GPT-56 Sol after effective spend roughly doubled: Astra can do novel things, but for coding it's tough to justify.
The initial excitement around Astra centers on totally new capabilities — video editing, 3D design and modeling — things that aren't currently part of most people's day-to-day work. Whether that's a good business strategy for OpenAI is a separate question.
Efficiency AI does your existing work better — faster, cheaper. Opportunity AI unlocks entirely new opportunities. That's usually been a mindset distinction rather than a difference in the models themselves; Astra is one of the first models where the difference is real.
There's nothing wrong with efficiency AI — it's the foundation of most AI use and most initial value. But companies that only think of AI as an efficiency technology will miss what the winners see: efficiency gains reset expectations for everyone, while the companies that chase new opportunities — even ones orthogonal to what they do today — race out ahead.
People don't walk around with a complete inventory of things they might make — we carry a smaller inventory shaped by our job, tools, and what people around us do. NLW's own opportunity use cases, like the pipeline that turns episodes into shareable chunks, came from stumbling, experimenting, and attacking existing problems in new ways.
Many of the 'opportunities' of opportunity AI aren't totally novel — they're things other people can already do that you couldn't. A good shorthand for stretching yourself is to look at what people in different roles are doing that you think is really cool, and bring it into your own work.
Marketing content has been visual, print, and video; AI opens up interactivity. Some of the most exciting first Astra experiments were building games — so why not games inside of work? Instead of telling someone a place rewards curiosity, give them a small mystery that makes them curious. A game implicates the audience's agency rather than treating them as a receiver.
A consultant whose 'cross-team decision-making' pitch sounds abstract could give a prospect a five-minute fictional product launch where sales promises a date, product finds a dependency, and support lacks information — letting the prospect encounter and get language for the exact problem the service solves. Other games can let customers exercise taste, like an awkward-apartment challenge built around your products.
Brands have experimented with games before, but they were constrained by development resources; now a solopreneur can experiment this weekend. Game design is still hard — far more games are released than ever become popular, even from professionals. Opportunity AI doesn't guarantee hits; it lowers the cost of the attempt.
The AI Daily Brief's clips aren't made by a dedicated clipping product like Opus — they come from a custom-built Claude-run pipeline that ingests the script and raw video and produces the output. These models may democratize video production the way coding agents democratized building software. The question: where could video help in your work — internal explainers, customer education, marketing?
Write a 60-second educational marketing script with AI, record it on your phone, then hand the video to Codex or Claude Code and ask for a visual motif, transitions, layered graphics, and a full production pipeline — so all you do is drop in source video. Give it one round of feedback, then ask: does this lower the barrier enough that video becomes a real tool for you?
A sales presentation has to choose one order; different buyers arrive with different questions. An exploratory demo organizes itself around actions — open this, isolate that component, inspect the result — so the visitor's question determines the path. Ask: what do customers need to inspect for themselves before your product makes sense?
Every proposal already contains an invisible model — assumptions about work, parallelism, resources, and timing. Instead of handing over one selected plan, expose parts of that model: 'we want it sooner' becomes 'we can finish sooner if the team attends more often.' Clients can explore trade-offs without every alternative becoming a new request for you to interpret.
Done well, this is a new way of interacting with clients that wasn't possible before — and it radically cuts the latency of back-and-forth in every negotiation. It has the feel of something that will become completely de rigueur, to the point we'll struggle to remember doing it any other way.
'We need another hire' hides several disagreements: process too slow, work unevenly distributed, an expected demand surge. People argue about conclusions while imagining different starting conditions. Building a simulator forces you to specify those conditions — and makes each assumption's consequences something you can inspect, like whether the decision holds if demand rises more slowly or training takes time.
A thread weaves through many of these ideas — interactive proposals, decision simulators — they're all what-if machines. What if we did this? What would the implications be? What trade-offs would it implicate? Making that reasoning manipulable is a core opportunity-AI move.
Some of the first genuine Astra excitement came from people driving 3D software like Blender — 3D walkthroughs of Zillow houses, learning experiences where rotating a digital object changes how you learn. 3D is a big part of what makes Astra unique and one of the best places to spend opportunity-AI experimentation time.
Organizations already have the raw material — interviews, project reviews, recordings, screenshots. An interview explains why a situation mattered; a recording shows what changed; a film combines them so the audience can recognize the problem, inspect the invention, and judge the outcome. First pass: dump the raw assets into Claude Code or Codex and let it one-shot the architecture just to feel the capability.
The hardest part of learning experiences is creating space for learners to exercise and get feedback on judgment — handling difficult customers, hard conversations with management. Simulation-style environments deliver immediate feedback on judgment midstream, and this will likely become an entire category of professional development experiences.
3D modeling makes physical products prototypable — an actual product, or something more abstract like a teacher's object with removable pieces that makes a difficult relationship tangible. The companion site's Make It Mine feature exists for exactly this: give it context about your work and let it surface possibilities, like acoustic panels shaped by a studio's own measurements.
Expert help is a conversation, but inside it you're gathering context, recognizing patterns, ruling out attractive-but-wrong options, and deciding what someone is ready for. Give a specific part of that judgment a form people can work through themselves — the interactive web app published alongside this episode is exactly that. It scales you, works for teammates as much as clients, and doubles as efficiency AI wherever you repeat the same things to different people.
The goal of opportunity AI is simply to stretch ourselves and ask what we might do now that we never would have before — several of these ideas turn out to be both at once. The point of the thought starters is that you don't have to know in advance what your opportunity looks like.