OpenAI released GPT-6 Astra on Thursday with a claim that no frontier lab has previously made in a product launch: that this represents the arrival of artificial general intelligence. Co-founder Greg Brockman stated at the launch event, "I do think we're there." The company's framing — "Welcome to the AGI era" — is explicit, not hedged.
The model's defining capability is computer use. Astra navigates browsers, fills CRM fields, operates desktop software including engineering tools like KiCad and FreeCAD, creates product listings from voice prompts, converts sketches into functional 3D models, and completes multi-step workflows across applications that have not been specifically integrated with any API. The benchmark scores are exceptional by any prior standard: 98.6 per cent on ARC-AGI-3, 97.6 per cent on FrontierMath Tier 4, 72.6 per cent on OSWorld 2.0 — completing tasks 47 per cent faster per workflow than its predecessor GPT-5.6 Sol. In cybersecurity testing, it reached OpenAI's internal "Critical" threshold and independently found two zero-day vulnerabilities.
Enterprise rollout begins September 5 via the Daybreak programme, followed by ChatGPT Plus, Pro, Business, and Enterprise, and availability on AWS Bedrock and Azure. API pricing is ten dollars per million input tokens and fifty dollars per million output tokens at standard rate; Fast mode doubles those figures. OpenAI's Strawberry programme for agentic enterprise customers has been briefing on Astra for several weeks in advance of the launch.
The computer-use architecture is the competitive differentiator that matters most for enterprise adoption. Previous AI models required custom API integrations with each tool a business uses — a significant deployment barrier that has slowed agentic AI adoption beyond the largest, most technically resourced organisations. Astra's ability to operate software through the visual interface, the way a human operator would, eliminates that barrier. Any enterprise workflow that a human performs on a screen is, in principle, within Astra's capability envelope. The practical implication for marketing operations, sales operations, finance, and any function that relies on desktop software is direct: the cost of automating routine workflow steps drops to the cost of an API call.
The commercial stakes are sharply illustrated by the demonstrated use cases. Creating eBay listings from a voice prompt. Filling Salesforce records from unstructured notes. Updating calendar entries across connected accounts. Operating engineering software to produce design files. These are not research demonstrations — they are the daily tasks of knowledge workers whose time is currently billed at hourly rates that no AI API cost structure approaches. If Astra performs these tasks reliably at scale, the economics of knowledge work change in ways that are difficult to model conservatively.
The missing number is more informative than the numbers present. OpenAI's own GDPval benchmark — introduced in 2025 to measure economically valuable real-world work, specifically designed to be the metric that would establish AGI by OpenAI's own standards — is absent from the Astra launch materials. The ARC-AGI-3 score deserves a footnote: the 98.6 per cent figure used a proprietary evaluation harness; separately, Nvidia achieved 100 per cent using a Claude Opus 5 agent operating the same benchmark, not Astra. The "Critical" cybersecurity threshold finding adds a layer of dual-use concern that Astra's restricted predecessors have already generated this week. None of these caveats diminish what Astra demonstrably does. They do suggest that the AGI claim is a commercial framing choice as much as a technical assessment, and that sceptical enterprise buyers will look for the GDPval number before updating their priors.
For MarTech practitioners, the immediate operational question is how Astra's computer-use capability affects the agency workflow stack. A model that can autonomously navigate a DSP, execute a creative brief in Canva, update a campaign tracker, and file a performance report — without API integration for any of those tools — compresses the billable workflow overhead that agencies and in-house teams currently absorb. The answer to how fast that compression happens depends on reliability at scale, which the first weeks of the Daybreak enterprise rollout will begin to establish.