OpenAI Just Made Your Worst Software Useful

OpenAI shipped GPT-6 Astra on Thursday.

Greg Brockman, the president of OpenAI, said this on release day: "Welcome to the AGI era."

Every AI newsletter in your inbox is going to spend 2,000 words arguing about whether that is true.

I am going to skip that fight. There is one line in the release that matters more to a law firm or a roofing company than the entire AGI debate, and almost nobody is talking about it.

Astra can operate your software by looking at the screen.

Not through an API. Not through a Zapier connection. Not through an integration your vendor promised last quarter and never built. It reads pixels, moves the mouse, types on the keyboard, and drives the interface the same way your intake coordinator does.

The thing that was actually blocking you

For three years the pitch to service businesses has been "AI will automate your back office."

Then you look at your actual stack. A practice management system built in 2011. A carrier portal that only works in one browser. A scheduling tool with no API. A CRM that technically has an API, except your vendor charges four figures a year to unlock it.

The intelligence was never the bottleneck. The plumbing was. Every automation project died in the same place: somebody had to build a connector, and the connector cost more than the problem.

Astra skips the connector. It scores 72.6% on OSWorld 2.0, the benchmark for driving a real computer, and OpenAI says it does that roughly 47% faster than the previous model. The tasks it was demoed on were form filling, CRM updates, calendar organization, and web research. That is the exact category of work you currently pay someone hourly to do.

The long tail of internal tools that never justified an integration just became addressable.

The rest of the numbers, quickly

• 96.0% on GPQA Diamond, graduate level science reasoning

• 98% on FrontierMath Tier 4, and 99.9% on ARC-AGI-3

• 57.9% on Terminal-Bench 4.0 for agentic coding, up from 37.3%

• Included in existing ChatGPT plans: Plus at $20, Pro at $100, Business at $200 per month

• API pricing at $10 per million input tokens and $50 per million output tokens

• Trained on more than 100,000 GPUs, and for the first time OpenAI used other AI models to supervise the training run

Now the part your competitors will skip

Astra hit the "Critical" threshold on OpenAI’s own cybersecurity framework. It scored 100% on ExploitBench and found two previously unknown vulnerabilities without human guidance. OpenAI restricted its most powerful cyber capabilities before shipping it.

OpenAI also acknowledged that Astra is harder to monitor in the evaluations designed to test whether a model can evade oversight. And nobody knows yet how it holds up on real work over months instead of benchmarks over minutes.

So here is the honest version: your integration backlog just turned into a governance backlog. The question stopped being "can it do the task" and became "who watches it, what credentials does it get, what gets logged, and what happens when the vendor changes their interface on a Tuesday."

If you hand an agent your CRM login and nobody reviews the output, you did not automate anything. You hired an employee with no manager.

What I would do this week

1. Write down the five tasks in your business where a human retypes information from one screen into another. That list is your automation queue, already in priority order.

2. Pick the one with the smallest blast radius. Not billing. Not client communication. Something like moving lead data between two systems, where a mistake costs you ten minutes instead of a client.

3. Give the agent its own login, never a shared one, and require a human to approve the output for the first thirty days. You are not testing the model. You are testing your own review process.

The businesses that win the next eighteen months are not the ones with the best model. Everyone has the same model. They are the ones who worked out which of their workflows an agent can safely run, and built the checks before they needed them.

Bottom line

Astra did not make AI smarter in a way you are going to feel day to day. It made AI able to use the software you already own, which is the thing you were actually waiting for whether you knew it or not.

Start with one workflow. Watch it for a month. Then add the second.

If you want help working out which workflows in your business are ready for this, hit reply and tell me what your team spends the most time retyping. I will tell you where I would start.

Dez

Co- Founder, Secret Agents AI