GPT-6 Astra First Impressions From Businesses

4 min video3 key momentsWatch original

⚡

TL;DR

Astra impresses business users with stronger computer use and reasoning abilities, solving previously unsolvable problems in media workflows, cost calculations, and code optimization.

Key Insights

Pause before executing — Astra pauses to understand new tasks before executing—a behavioral shift that produces measurably better results than models that jump straight in.

First time solving it right — Box solved a media cost-calculation problem for the first time with Astra by having it verify assumptions and catch double-counting errors a human would miss.

3.3% at scale matters — A 3.3% GPU efficiency improvement on workloads running across thousands of GPUs compounds into massive time savings—Astra found it by testing and iterating on past experiments.

Want this for every new video OpenAI posts? Brevyd summarizes each upload automatically, the morning it drops.

Deep Dive

The coworker feeling

Multiple business users describe Astra as unusually human-like in how it approaches tasks. One executive calls out the model's confidence and deliberate pacing—it takes time to understand a task before diving in, which sets it apart from prior models. Another frames it as a real coworker, largely because of how well it handles computer use. The common thread: this model raises ambition levels and produces results on research problems teams hadn't been able to crack before.

Media workflows and cost calculations

Box highlighted two use cases. In media and entertainment, Astra handles calculating incentives and verifying assumptions with enough thoroughness that it catches double-counting errors in cost models—something previous models failed at. A creative team also showed Astra generating a static vector animation in Flora, a node-based editing tool, without consulting documentation. The model just used computer use and browser navigation to build workflows, pull nodes, and apply image generation. One executive said watching it work felt magical.

Research and infrastructure optimization

Another company is using Astra to steer smaller, cheaper models toward specific signals in raw text. Astra excels at balancing breadth and depth, exploring each data node thoroughly while validating assumptions—something earlier generations struggled with. The standout result: the model found a 3.3% efficiency gain on a workload running across thousands of GPUs. That sounds marginal until you run it at that scale. The ability to parallelize testing and iteration means shipping product faster and accelerating workflows across infrastructure.

Takeaways

✓Use Astra's verification instinct on high-stakes calculations where double-counting or assumption drift could cascade—media costs, revenue splits, budget forecasts.
✓Let Astra explore documentation-free workflows in visual tools where you'd normally need a human to navigate UI and context.
✓Feed Astra past experiments and logs when optimizing workloads; its assumption-checking catches edge cases that incremental tuning misses.

Key moments

0:19Astra's confidence in context

“Its ability to really give itself time to understand a new task before it actually dives in. Something that makes it feel like it leads to far better results down the line.”

1:27First time solving media math

“Imagine if it did that wrong, it would have double counted this 20% incentive and gotten your entire cost wrong. Is it the very first time you're seeing a model get this right? Yes, this is one of the first times.”

3:27GPU efficiency breakthrough

“Finding like new optimization that made it 3.3% faster, which is maybe doesn't sound like a lot, but it's actually like really significant for like a workload that you run on like thousands of GPUs.”

You just read one. Brevyd does this for every upload.

Follow OpenAI and every new video comes back as a summary like this, in your morning briefing. No watching required.