OpenAI released a new AI model called GPT-6 Astra on September 3. Developers with early access spent the following two days testing it across coding, art, music and computer tasks.
The model costs $10 per million input tokens and $50 per million output tokens. A token is roughly three quarters of a word. That price is 2.5 times higher than GPT-5.6 Sol, the model it replaces.
OpenAI president Greg Brockman used the launch event to say the model represents the arrival of AGI, a term used for AI systems that can match human ability across a wide range of tasks.
Computer Use and Cybersecurity
The main new feature is called computer use. This lets the model move a mouse and type on a keyboard on a real desktop instead of just giving instructions for a person to follow.
On a test called OSWorld 2.0, which measures how many desktop tasks an AI can finish without help, Astra scored 72.6%. That test took about 40 minutes per task. Sol scored 65.7% and took about 75 minutes.
OpenAI also gave Astra its highest rating yet for cybersecurity risk. This means the model can locate unknown flaws in software and build a working attack without a person pointing it in the right direction first.
What Early Testers Built
Investor Matt Shumer spent a week letting Astra build a copy of Manhattan inside Unreal Engine, the software used to make the game Fortnite. He said the model worked block by block until each street looked right.
Shumer also asked the model to create a survival game world filled with characters, each one run by a separate copy of Astra. He left it running overnight and returned to find the characters talking to each other on their own.
Developer Tom Krcha gave Astra one photo of a house and received back a full working 3D version of the interior, including appliances and toys, running at 60 frames per second.
A developer using the name SuSu had Astra rebuild the Chinese city of Hangzhou in a web browser in 24 minutes. The result included landmarks, a day and night switch and clickable buildings.
Games came up often in the tests. Former Apple designer Anshu Chimala said Astra built a full 3D game in 45 minutes using only a small part of his usage limit.
Developer Rishi Prasad built a browser based multiplayer shooter in a single day, complete with voice chat and support for 12 players at once.
Musician and blogger Auggie tested Astra with a request to write a four part chorale in the style of composer Johann Sebastian Bach. The model produced the highest scoring result the test has recorded, with no rule breaking errors in the melody.
The results were less positive for writing tasks. Editor Louis-François Bouchard ranked Astra 11th out of the models he tests for writing quality, while the older Sol model ranked 6th.
Independent testing firm Artificial Analysis recorded a drop of about 80 points on a benchmark that scores AI performance on real world professional tasks across 44 jobs.
Author Giuseppe Paleologo asked Astra to generate new ideas about investment portfolio strategy. He said the results were a mix of common knowledge and text that read as clearly written by a machine.
Not every review was negative. A staff writer at the outlet Every let Astra draft the first version of an article, and the company's CEO read it without noticing it had been written by AI.
OpenAI has not announced a date for expanding access to Astra beyond the early testers who received it during the September 3 launch window.
