.png){kind=logo}
GPT-6 Astra scores about the same as its predecessor on general intelligence tests but leads sharply on tasks that require operating software.
Coding gains come from cost and speed on long agent sessions, not from writing better code in a single response.
OpenAI has classified Astra at a new critical risk level for cybersecurity, adding fresh weight to how companies manage its access.
OpenAI released GPT-6 Astra in September 2026, and the launch pitch was simple: this is the model that finally uses a computer the way a person does. That framing is worth pausing on. Most model launches promise sharper answers. Astra's real pitch is different. It promises fewer humans standing between an instruction and a finished task. That distinction matters more than it sounds.
A model that answers well still needs someone to copy the answer, open the right app, click the right button, and check the result. A model built for execution is meant to do all of that on its own. Astra is OpenAI's clearest attempt yet at closing that gap, and the results are mixed in a way that is more interesting than a simple upgrade story.
Independent testing from Artificial Analysis found Astra's overall intelligence score nearly identical to GPT-5.6 Sol, its predecessor, and trailing some rival models on general reasoning. That is not a criticism buried in the data. It is the headline. Astra was not built to think harder. It was built to carry out longer sequences of actions without losing track of the goal.
The clearest proof lies within OSWorld 2.0, a test that measures whether a model can complete real desktop tasks such as opening files and navigating apps. OpenAI reports Astra scoring around 73 %, compared with about 66 % for Sol, and finishing close to half the time. Independent write-ups from DataCamp and MindStudio landed in a similar range.
On ScreenSpot-Pro, a test of locating and clicking screen elements without external help, both OpenAI and outside reviewers placed Astra in the low nineties, well past Sol's high seventies.
On common coding benchmarks like DeepSWE and Frontier Code, MindStudio found Astra running close to even with Sol and with rival models. That is a tie, not a win. Where Astra pulls ahead is Terminal-Bench 4.0, a test built around long, messy terminal sessions that require chaining many commands and recovering from mistakes along the way.
Artificial Analysis found Astra matching a top rival coding model on its Coding Agent Index while running under half the cost per finished task. The reason comes down to tokens. Astra tends to finish long coding sessions using far fewer of them.
So the upgrade is not sharper code from a single prompt. It is lower cost and faster completion across sessions that stretch for hours, which describes most real engineering work far better than a short coding quiz does.
One vivid example came from Matt Shumer, former chief executive of HyperWrite, who let Astra run inside Unreal Engine, the game engine behind Fortnite. Over an extended session, it built a working three-dimensional replica of part of Manhattan. That kind of result depends on holding spatial context across hundreds of linked actions, not producing one clever line of code.
Also Read: 5 Ways OpenAI’s NextSlide Acquisition Could Transform ChatGPT
On benchmarks built around real office and technical work, Astra posted its strongest scores. On Agents' Last Exam, which spans financial modeling, engineering, and media production inside real software, it edged past both Sol and a leading rival reasoning model, using notably fewer tokens along the way.
What matters here is not that a model can build a spreadsheet. Older models already managed that with a person guiding each step. The shift is that Astra can move between research, calculation, software interaction, and final output with less manual handoff between stages. That turns AI from a drafting assistant into something closer to a task owner.
Also Read: OpenAI Brings ChatGPT Ads to India for Free, Go Users
Astra costs about two and a half times more per token than Sol. For short, simple requests, that premium rarely pays off. For long agentic sessions involving hundreds of tool calls, lower token use can offset or beat the higher price on a per-task basis. The economics favor Astra as task length grows, not as a blanket rule.
OpenAI also confirmed Astra is its first model to reach a critical risk level for cybersecurity capability, meaning it can find unknown software flaws and build working exploits with the right access. New safeguards followed, including tighter review of high-risk actions inside ChatGPT and Codex.
For any company weighing 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.
1. What is the GPT-6 Astra?
GPT-6 Astra is OpenAI's agentic AI model designed to go beyond generating responses by operating computers, using tools, coding, conducting research, and completing multi-step professional tasks.
2. Is GPT-6 Astra better at coding?
GPT-6 Astra's biggest coding advantage is in long, tool-heavy workflows rather than basic code generation. It can handle extended terminal sessions, recover from errors, and complete complex development tasks with fewer tokens.
3. How does GPT-6 Astra differ from GPT-5.6 Sol?
Astra's aggregate intelligence is broadly similar to GPT-5.6 Sol, but it is stronger at agentic execution. Its advantages are most apparent when tasks require computer interaction, sustained tool use, and multiple sequential actions.
4. Is the GPT-6 Astra expensive to use?
GPT-6 Astra has a higher per-token cost than Sol, making it less economical for simple requests. Its economics become more attractive for long agentic workflows where lower token consumption and reduced human intervention can lower the total cost of completing a task.
5. What are the cybersecurity risks of GPT-6 Astra?
Astra has reached OpenAI's critical cybersecurity capability threshold, meaning it can potentially discover software vulnerabilities and develop exploits when given appropriate access. Organizations therefore need strict permissions, monitoring, approval controls, and audit mechanisms.
