My AI Might Actually Do My Errands Soon, and I'm Listening
OpenAI's new 'errand-running' benchmark suggests future AI like GPT-6 Astra could finally become practical partners for daily tasks like shopping and planning.
The short version
OpenAI is now measuring how well its AI can do real-life tasks like shopping or planning a meal. Their newest model, GPT-6 Astra, scored an 80% on this new benchmark, a huge leap from older versions. This is one of the first real signs that we’re getting AI agents that can actually help with our day-to-day lives, not just write emails.
What did OpenAI just announce?
I’ve been waiting for this. For years, AI has been about abstract power. Now, it might be getting practical. OpenAI shared that it has created a new way to measure how well its models perform on everyday human tasks. Think of it as a benchmark for running errands. The kind of stuff we all do. Like creating a weekly budget, planning meals, or figuring out an online shopping order. According to OpenAI’s report on AI errands, this is a major focus for them now.
The company is putting its new models to the test on these real-world scenarios. The results are what caught my eye. OpenAI claims its latest model, which they’re calling GPT-6 Astra, scored 80% on general performance across these tasks. That’s a massive jump. A really big one.
To put that number in context, they also shared a score for a previous model, GPT-5.2. Back in December 2025, that model only managed a 49.3%. Going from a coin-flip chance of success to a solid B- grade in less than a year is a significant acceleration. It suggests the underlying technology is getting much better at understanding and executing complex, multi-step instructions that mimic how people actually think and plan.
Why is this different from other AI tests?
This matters because most AI benchmarks are deeply academic. They are. They feel like a final exam for a PhD program I never enrolled in. You hear about things like MMLU, which tests massive multitask language understanding, or the ARC Challenge, which tests grade-school science knowledge. These are important for researchers. They push the science forward. But a perfect score on a test of graduate-level physics questions doesn’t tell me if the AI can help me figure out a vacation itinerary for my family of four that keeps everyone happy and doesn’t break the bank.
This new benchmark from OpenAI is different. It’s measuring something else entirely. It’s measuring usefulness.
The shift is from testing passive knowledge to testing active capability. It’s the difference between an encyclopedia and an assistant. An encyclopedia knows a lot of things. An assistant does things. This new evaluation framework is explicitly designed to measure the model’s ability to act as an agent—to take a high-level goal and break it down into smaller, actionable steps, and then execute them in a logical order. That’s a whole new ballgame.
For example, the old way was asking an AI, “What are the symptoms of diabetes?” The new way is, “Create a seven-day meal plan for my newly diagnosed diabetic father, ensuring it’s low-glycemic, uses ingredients we can buy at Trader Joe’s, and includes at least one meal he actually enjoys, like lasagna.” One is a lookup. The other is a project. OpenAI is finally starting to measure the project.
How much better is this new AI, really?
Let’s talk about the numbers, because they tell a story. The jump from GPT-5.2’s 49.3% to GPT-6 Astra’s 80% is the headline. A model that gets things right less than half the time is a toy. It’s frustrating. You’d spend more time double-checking and correcting its mistakes than you’d save by using it in the first place. I’ve been there. We’ve all been there, trying to get an AI to do something simple and watching it fail in a uniquely spectacular way.
But 80%? That starts to cross a critical threshold. It moves from unreliable novelty to a potentially dependable tool. It’s not perfect. A 20% failure rate still means it gets one out of five tasks wrong. You’re not going to hand over your credit card and say “book my flights” just yet. But for low-stakes planning? For generating a first draft of a grocery list or outlining a project plan? An 80% success rate is something you can work with. It’s a capable intern.
The even more impressive number is the 90.4% score OpenAI reported for a very specific task: the one about creating a meal plan for a diabetic parent. The complexity here is what’s important. This task has multiple constraints. Medical needs (low-glycemic). Logistical needs (ingredients from a specific store). And even emotional needs (a meal the person will actually enjoy). Hitting over 90% on something that nuanced suggests the model is becoming adept at handling the messy, overlapping requirements of real life. That’s the part that feels like science fiction bleeding into reality.
So, should you be expecting an AI helper soon?
No. Not tomorrow, anyway. It’s important to keep our feet on the ground here. This news from OpenAI is a peek into the lab, not a product launch. GPT-6 Astra is an internal research model. This errand-running benchmark is an internal tool. You can’t download this. You can’t use it. It’s a signpost pointing to the future, not a vehicle that’s ready to drive off the lot.
There is a huge gulf between a model achieving a score in a controlled test and a product that works reliably for millions of people in the wild. What happens when the AI tries to access a website and it’s down? How does it handle an out-of-stock item on your shopping list? How does it ask for clarification when a request is ambiguous? These are the unglamorous, difficult engineering problems that turn a powerful model into a useful product.
Furthermore, an AI agent needs to connect to the real world to be truly useful. It needs access to your calendar, your email, your bank accounts, and shopping websites. This opens up a giant can of worms around privacy, security, and trust. Giving an AI the ability to do things on your behalf is a much bigger step than giving it the ability to say things. We’re still in the very early days of figuring out how to build those connections safely and effectively.
Think of this announcement as a very exciting concept car. It shows what might be possible one day. But it will be a while before that technology shows up in the car you and I can actually drive.
What’s the catch with this announcement?
With any big announcement from an AI lab, you have to ask what they aren’t saying. There are a few things to keep in mind here. First, we are taking OpenAI’s word for all of this. The benchmark was created by OpenAI, the models were built by OpenAI, and the scoring was done by OpenAI. It’s not an independent, third-party audit. It’s a self-reported grade. While the numbers are impressive, it’s like a student grading their own homework. Until others can replicate and verify these results, a healthy dose of skepticism is warranted.
Second, the methodology is a black box. What does “80% performance” actually mean? How are the tasks graded? Is it a human looking at the result and giving a thumbs-up? Is it another AI model making the judgment? What constitutes a pass versus a fail? A plan that’s 95% perfect might still be a failure if the 5% error is putting a peanut-based sauce on the plate of someone with a deadly allergy. The details of the evaluation rubric are everything, and we don’t have them.
Finally, this is a strategic move in a fiercely competitive market. OpenAI is in a constant battle for talent, investment, and public perception against rivals like Google, Meta, and Anthropic. Releasing these kinds of impressive-sounding results helps them control the narrative. It frames the conversation around a metric they’ve just defined and, conveniently, are winning at. It’s as much a public relations maneuver as it is a scientific disclosure. It’s exciting, yes. But it’s not purely academic.
FAQ
What is GPT-6 Astra? GPT-6 Astra is the latest internal AI model from OpenAI. The company claims it shows significant performance improvements on practical, everyday tasks, but it is not yet available to the public.
Can I use this new AI errand-runner today? No. The announcement is based on an internal benchmark and research model. It indicates the future direction of OpenAI’s work but is not a product or feature you can use right now.
How is this different from ChatGPT? ChatGPT excels at conversational tasks like answering questions or generating text. This new research is focused on measuring an AI’s ability to be an “agent”—to understand and complete complex, multi-step goals like planning a trip or creating a budget, which requires a deeper level of planning and execution.
Is OpenAI’s 80% score trustworthy? It’s a company-reported score on its own internal test, so it has not been independently verified. You should view it as a promising milestone that OpenAI is sharing, but not as a definitive, objective measure of its capabilities just yet.