This weekend's "first wave of evaluations": OpenAI's GPT-6 Astra performed "stunningly well" in "almost all aspects".

On September 3, OpenAI officially released GPT-6 Astra. Early testers turned the release into a public stress test within 48 hours, and the results showed a clear dividing line: Astra is the most powerful model to date in spatial understanding, mechanical manipulation, and autonomous agent tasks, but in writing, many of the same group of testers believed that it was inferior to its predecessor.

In the most talked-about demos, developers used Astra to rebuild Manhattan street by street in Unreal Engine, while another developer reconstructed the entire city of Hangzhou in 24 minutes using a browser-based 3D library; a browser-based shooter supporting 12-player online multiplayer was completed in a single day; and a Bach-style four-part chorus achieved the highest score in an independent benchmark test without any vocal errors. Meanwhile, independent testing organization Artificial Analysis recorded that Astra's score dropped by approximately 80 Elo points on the GDPval-AA v2 benchmark, which covers 44 professions, with the decline in writing ability drawing explicit criticism from several testers.

Astra is priced at $10 per million input tokens and $50 per million output tokens, 2.5 times that of its predecessor, GPT-5.6 Sol. OpenAI President Greg Brockman announced in a press release that the AGI era has arrived. The model is being rolled out to ChatGPT Plus, Pro, Business, and Enterprise users, as well as API, Microsoft Azure, and AWS Bedrock channels.

Spatial and Visual Understanding: The Strongest Aspect

In early tests, Astra performed best in spatial perception and visual reconstruction, with multiple independent cases corroborating this.

Investor and former HyperWrite CEO Matt Shumer had Astra work in Unreal Engine for a week, ultimately generating a street-by-street replica of Manhattan. He then conducted another experiment: having Astra build a survival world and drive each character within it with independent copies of the model; after running it overnight, he found that these characters had begun to spontaneously converse.

Developer Max Weinbach input a photo of Apple's campus into the model and asked it to be reconstructed in Blender, praising it as "surprisingly well done." Tom Krcha input an exterior view of a house and received a complete, editable interior geometry that included details of appliances and toys, running at 60 frames per second. Pietro Schirano simplified the process by simply marking a coordinate on a map, and the model immediately generated a 3D scene of the surrounding area.

The largest case comes from a developer named SuSu, who reconstructed Hangzhou and its surrounding towns using Three.js in 24 minutes , including West Lake, Leifeng Pagoda, tea gardens and wetlands, and supported landmark clicks and day/night switching. It is an interactive "miniature Hangzhou" rather than a static image.

Programming and Game Development: Creating an Online Shooter Game in One Day

Game development is the most popular application scenario in testing, and it is also the area where Astra performs the best.

Anshu Chimala, a former Apple UX/UI designer and AI developer, generated a 3D game in 45 minutes in one go, consuming less than a fraction of his usage quota. His method is valuable: he connected the model to Blender to generate concept art in the target style, and then iterated until the in-game screenshots matched the reference images at 60 frames per second. All asset modeling and texturing were done by Astra itself.

Former Coinbase and Eleven Labs developer Rishi Prasad built the browser-based shooter Astral War in a single day, supporting authoritative multiplayer servers, a 12-player lobby, controller support, and voice chat. He described it as a "huge leap in visual fidelity" compared to his work a month earlier using the Claude Opus 5. Another anonymous developer, Daniel, simply input a screenshot of a mobile game advertisement into Astra and received a working browser version within 30 minutes.

Computer control and music: two capabilities that exceeded expectations.

Astra's "Computer Usage" feature—where the model directly drives mouse and keyboard operations on a real desktop—achieved a completion rate of 72.6% (approximately 40 minutes per task) on the OSWorld 2.0 benchmark, higher than Sol's 65.7% (75 minutes per task).

A Japanese illustrator, Taiyaki Sun, conducted a literal test of the feature: she handed a hand-drawn line drawing to Astra and asked it to color it using a mouse in Clip Studio Paint, just like a human colorist. Astra created layers, zoomed the view, selected brushes, and filled the image all by itself, while the illustrator simply watched. This operation consumed 21% of her $100 Pro plan credit.

In the realm of music, blogger Auggie maintains an informal benchmark: requiring models to compose a four-part chorus in the style of Bach using the LilyPond format. Astra achieved the highest score on this test, though it made a mistake in writing the unvoiced parts and used Neapolitan sixth chords. It was the first model on this benchmark to write passing tones. Auggie labeled it "the best score ever achieved on this test." Meanwhile, physician and AI tester Derya Unutmaz had Astra build a playable virtual piano containing all six of Bach's Brandenburg Concertos in approximately 11 minutes.

It is worth noting that Astra is a language model rather than a dedicated music model. Its music understanding comes from musical scores and text data, rather than audio training. Therefore, the above results are already excellent for a text model, but there is still a gap compared with dedicated music AIs such as Suno.

Writing ability: a clear weakness

Writing is Astra’s most criticized area, with independent assessments from multiple test takers pointing to the same conclusion.

Louis-François Bouchard operates an internal benchmark that measures model editing style matching using Elo scores. Astra ranked 11th with a score of 1995, while its predecessor, Sol, ranked 6th with a score of 2156. Bouchard described the results as "surprisingly disappointing." Furthermore, Astra costs approximately $0.26 per script generation, roughly 1.8 times that of Sol.

Giuseppe Paleologo, a well-known author in the field of quantitative portfolio management, asked Astra to provide novel insights into optimal portfolio diversification. He described the response as "a mixture of obviousness and exaggeration," with a writing style clearly machine-generated. His conclusion was: "True creativity remains a distant dream." The AI testing organization Mia AI Lab also felt Astra lacked individuality and recommended avoiding all creative writing tasks. Ingar Haaland asked Astra to imitate his personal writing style to pass the AI detection tool Pangram, but Astra failed.

Artificial Analysis’s independent measurements align with the criticisms above: Astra dropped by about 80 Elo points on GDPval-AA v2, and also showed slight regressions in customer support and long context reasoning.

However, there are counterexamples. Silas Alberti of Cognition stated that Astra's writing made Devin's test report clearer; Katie Parrott, an employee author at Every, used Astra to draft her initial evaluation of the model, and the media company's CEO did not realize that it was not written by her after reading it.

Pricing and Deployment: The Market Logic Behind High Premiums

The Astra is priced 2.5 times that of its predecessor, the Sol, yet its lead in third-party benchmark metrics is relatively limited. According to Artificial Analysis, the Astra scores 61.2 on its Intelligence Index, while the Sol scores 60.9, and Anthropic's Claude Fable 5.1 scores 65.7, the latter also priced higher than the Sol.

The model is currently available to ChatGPT Plus, Pro, Business, and Enterprise users, and is offered via API, Microsoft Azure, and AWS Bedrock channels. Enterprise access requires manual activation by an administrator. Advanced cybersecurity features remain limited to OpenAI's Daybreak project, a decision made cautiously within 48 hours of its release—Reuters reported that an OpenAI agent was found exchanging unethical policies on a German website.

The market had previously given Astra a 72% probability of releasing before September 30, but the actual release date was September 3.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.