GPT-6 Astra is out - what OpenAI's agent model changes for business automation
GPT-6 Astra is built to run a computer on its own, doing real desktop tasks in about 40 minutes instead of 75. What it changes for automation, and what it does not.
OpenAI released GPT-6 Astra on September 3. It is available now in ChatGPT paid plans, the OpenAI API, Microsoft Foundry, Amazon Bedrock, and GitHub Copilot. Standard API pricing is $10 per million input tokens and $50 per million output tokens, with a fast tier at $20 and $100. Prompts over 272,000 input tokens are billed at double the input rate.
The pitch is agents. OpenAI is positioning Astra as a model built to operate a computer on its own: open a browser, fill in forms, update records in a CRM, work through a spreadsheet, run multi-step research, with less human correction along the way. Greg Brockman used the launch to say OpenAI now considers itself in "the AGI era." Set that aside. The benchmark numbers are the useful part.
What is genuinely new
On OSWorld 2.0, a test of real desktop tasks, OpenAI reports Astra scoring 72.6% on an offline subset, up from 65.7% for GPT-5.6 Sol. It also reports Astra finishing those tasks in about 40 minutes instead of about 75. The second number matters more than the first. A computer-use agent that needs 75 minutes per task and gets two-thirds of them right is a demo. Cut the time nearly in half and raise the hit rate, and it starts to be something you can put behind a work queue.
The academic scores are high. VentureBeat lists 96% on GPQA Diamond, 97.6% on FrontierMath Tier 4, and 74.1% on DeepSWE. Those numbers have been climbing all year, and most business work does not look like a math contest. Treat them as table stakes, not as the reason to switch.
OpenAI also says Astra is the first model it has rated "Critical" for cybersecurity capability under its own Preparedness Framework. In plain terms, it can find unknown software vulnerabilities and build working exploits without a person directing each step. That is not a capability you will use. It is a signal about where the line now sits, and why access to the top tier comes with more gating and monitoring attached.
What it means if you run a business
The workflows Astra is sold on are the ones that have resisted automation because they need a screen and a judgment call. Pulling data out of a vendor portal with no API. Reconciling two systems that do not talk to each other. Processing an inbox of non-standard requests. Keeping a CRM current from email threads. If you scoped one of these last year and gave up because the model wandered off task after ten minutes, it is worth testing again now.
Pricing is where the framing gets slippery. OpenAI says to think in price per task, not price per token. For agent work there is something to that. A task that finishes in 40 minutes with fewer retries can cost less than a cheaper model that loops for an hour. But that only holds if you measure it. Run your actual workflow, count the tokens and the failures, and compare against what you pay today. Watch the 272,000-token pricing step: agents that carry long histories or drop whole documents into context can cross it without anyone noticing, and the input bill doubles when they do.
The honest caveat
OpenAI did not publish GDPval results for Astra. GDPval is its own benchmark for real-world economic tasks across 44 occupations, and it was built to measure the exact claim this launch is making. Leaving it out of the announcement is a gap worth noticing.
More capable agents are also harder to supervise, not easier. OpenAI says its internal testing showed Astra never exceeded its authorized scope, where an earlier model did so 48% of the time without safeguards. That is OpenAI testing OpenAI, not independent verification. A model that does more with less visible reasoning is one you need to wrap in tighter permissions, logging, and review. The safety machinery cuts both ways too: OpenAI notes that legitimate workflows can be slowed or stopped outright by its safeguards, which is its own operational risk to plan for.
None of this changes the basic rule. A model that is better at computer use makes a well-defined task cheaper to automate. It does not define the task for you, and it does not fix a process that is broken on paper.
What to do about it
Pick one workflow you have already tried to automate and abandoned because the agent was not reliable enough. Re-run it on Astra with the same prompts. Log every step, and measure cost and failure rate against your earlier attempt. If it clears your bar, expand from there. If it does not, you have spent an afternoon and you know where the line still is.
Want this kind of system in your business? Book a free scoping call.