OpenAI launched GPT-6 Astra in September 2026, but the message was not about smarter answers. The company framed it as the first model that uses a computer the way a person does: opening files, navigating apps, and completing multi-step tasks without constant human hand-holding. That is a meaningful shift in what an AI model is expected to do.
Most model launches focus on reasoning benchmarks. Astra breaks that pattern. It was not built to think harder; it was built to execute longer sequences of actions. Independent testing from Artificial Analysis found Astra’s general intelligence score nearly identical to its predecessor, GPT-5.6 Sol, and trailing some rivals on pure reasoning. Yet on execution-focused tests like OSWorld 2.0, which measures desktop task completion, Astra scored around 73 percent compared with Sol’s 66 percent, and finished tasks in roughly half the time. On ScreenSpot-Pro, a test of clicking screen elements without help, Astra reached the low nineties, well above Sol’s high seventies.
Coding benchmarks tell a similar story. On standard tests like DeepSWE and Frontier Code, Astra roughly ties Sol and rival models. But on Terminal-Bench 4.0, which simulates long, messy terminal sessions with chained commands and error recovery, Astra pulls ahead. Artificial Analysis reports that Astra matches a top rival coding model on its Coding Agent Index while costing less than half per finished task. The advantage is not better code from a single prompt; it is lower token usage across sessions that stretch for hours—the kind of work real engineers do.
A vivid demonstration came from Matt Shumer, former CEO of HyperWrite, who let Astra run inside Unreal Engine, the game engine behind Fortnite. Over an extended session, the model built a working 3D replica of part of Manhattan. That required holding spatial context across hundreds of linked actions, not producing one clever line of code.
Professional task benchmarks also show Astra’s strengths. On Agents’ Last Exam, which covers financial modeling, engineering, and media production inside real software, Astra edged past both Sol and a leading rival reasoning model while using fewer tokens. The key is not that a model can build a spreadsheet—older models could do that with human guidance. The shift is that Astra can move between research, calculation, software interaction, and final output with less manual handoff between stages. That is a turn from drafting assistant toward task owner.
Cost and risk remain central considerations. Astra’s per-token price is about 2.5 times higher than Sol’s. For simple requests, that premium rarely pays off. For long agentic workflows involving hundreds of tool calls, lower token consumption can offset the higher price and even beat Sol on a per-task basis. The economics favor Astra as task length grows, not as a blanket rule.
OpenAI also confirmed that Astra is its first model to reach a critical cybersecurity risk level. It can find unknown software flaws and build working exploits with the right access. The company introduced new safeguards, including tighter review of high-risk actions inside ChatGPT and Codex. For any organization considering deployment, the real question is not whether Astra performs well—it is how much authority any AI agent should hold before permissions, audit logs, and rollback controls become non-negotiable.
The next race among AI labs will not be won on cleverness alone. It will be won on which system can be trusted to act without supervision and still be pulled back safely when something goes wrong.
