I'm working on a website build right now.
Then I saw OpenAI's announcement that ChatGPT Work is now powered by GPT-6 Astra: https://lnkd.in/dBRR8jbr
So I tested it immediately on a real task: a full technical audit of the site.
The result was genuinely solid. The agent opened the homepage, two case-study pages and seven linked documents. It checked the page structure, links, contact form, metadata, contrast and several other details. In effect, it completed a small technical audit on its own — the kind of task where a person would spend considerable time simply navigating between pages, files and checks.
The agent paid for that efficiency with my usage limit. 7 minutes 36 seconds of work — and 80% of my five-hour usage window was gone. I'm currently on the ChatGPT Plus plan.
One detail matters here: this allowance is not specific to Work mode. According to the usage screen, it is shared across Codex, Work, Workspace Agents and ChatGPT for Excel. One budget for several agentic tools, not a separate allowance for each type of work.
This brought me back to a question I raised under another OpenAI post a few days ago, about the economics of data agents when GPU fleet utilisation is already above 90%. I don't know whether these two things are directly connected. But the experiment suggests that the practical constraint on scaling agentic work may no longer be only what the model can do. It may also be how much compute the product can make available for a single real-world task — and how quickly a user's allowance is consumed.
The agent was efficient at solving the task. Whether that efficiency is broadly accessible may be the more interesting question.
#AI #AgenticAI #GPT6Astra #ChatGPTWork