OpenAI released GPT-6 Astra on September 3, pricing the model at $10 per million input tokens and $50 per million output tokens—2.5 times the rate of GPT-5.6 Sol. Within 48 hours, developers with early access demonstrated a stark division in the model's capabilities: exceptional performance in spatial, mechanical, and coding tasks, but measurable weakness in creative writing.
Spatial and Visual Understanding
Early testers showcased Astra's spatial abilities across multiple projects. Matt Shumer used the model to generate a street-by-street replica of Manhattan in Unreal Engine over the course of a week. Max Weinbach fed it photographs of Apple Park and requested a reconstruction in Blender, describing the results as executing "an absurd job." A developer using the handle SuSu rebuilt the Chinese city of Hangzhou and surrounding towns in Three.js within 24 minutes, creating an interactive web-based model complete with landmarks and a day-night toggle.
Game Development
Game creation emerged as a primary use case. Anshu Chimala, a former Apple designer, generated a complete 3D game in 45 minutes using a workflow that connected the model to Blender for iterative asset generation and texturing. Rishi Prasad, a former developer at Coinbase and Eleven Labs, built Astral War—a browser-based multiplayer shooter with 12-person lobbies, controller support, and voice chat—in a single day. He characterized the results as "a huge, step-function leap in visual fidelity" compared to what he achieved with Claude Opus 5 a month earlier.
Music Composition
Astra posted the strongest result on an informal Bach benchmark maintained by Auggie, an AI tester. The model generated a four-part chorale in G minor with no voice-leading errors and a Neapolitan sixth chord, becoming the first model to include passing tones on this test. On OpenScore String Quartets, which measures accuracy in reading and transcribing classical scores, Astra reached 0.84 compared to 0.19 for Sol.
Computer Use Capability
The model demonstrated the ability to control a mouse and keyboard on a desktop environment. On OSWorld 2.0, a test measuring autonomous completion of ordinary desktop tasks, OpenAI reported 72.6% success at roughly 40 minutes per task, compared to 65.7% at 75 minutes for Sol. A Japanese illustrator had Astra color hand-drawn line art in Clip Studio Paint, with the model managing layer creation, brush selection, and zooming independently.
Writing and Creative Work
Multiple independent assessments documented a decline in writing quality. Testers rated Astra's writing below its predecessor. Artificial Analysis measured a drop of roughly 80 Elo points on GDPval-AA v2, a benchmark covering economically valuable tasks across 44 occupations. Ingar Haaland asked Astra to write four paragraphs in his own style closely enough to evade Pangram, an AI-detection tool. The test failed: "Pangram is not fooled." Mia AI Lab characterized the model as lacking personality and advised against using it for creative work.
Pricing and Access
Astra is rolling out to ChatGPT Plus, Pro, Business and Enterprise users, through the API, and via Microsoft Azure and AWS Bedrock. Advanced cybersecurity features remain gated behind OpenAI's Daybreak program. On the Artificial Analysis Intelligence Index, Astra scores 61.2 against 60.9 for GPT-5.6 Sol and 65.7 for Anthropic's Claude Fable 5.1.
