The arrival of the Astra model to individual subscribers marks a historic shift, making direct control of software interfaces and web browsing available to the wider public for the first time.
Article Index:
- Astra begins rolling out to individual tiers
- Exceptional software control capabilities
- Benchmark results and risk classification
- Staged deployment and pricing policy
- Mixed reactions and creator community criticism
- Frequently asked questions
Astra begins rolling out to individual tiers
OpenAI has started rolling out its newest and most advanced model in history, “GPT-6 Astra”, to ChatGPT Plus paid subscribers for $20 a month, following its limited initial launch on September 3 to select organizations and higher subscription tiers. This model represents a complete leap in releases compared to the “GPT-5.6 Sol” model, and also inaugurates a brand-new phase in AI-driven computer use.
OpenAI describes the Astra model as “the smartest and most compatible model in the world,” noting that making it available to individual users at standard subscription prices reflects its commitment to delivering advanced computing capabilities to a broad audience of developers and everyday users rather than limiting it solely to major financial and research institutions.
Exceptional software control capabilities
The “computer use” feature heads up the model’s superpower interface, enabling it to browse the internet, use various software and application programs, and manage complex computing operations on the user’s screen without needing integration via custom application programming interfaces (APIs). Greg Brockman, co-founder of OpenAI, stated in a press briefing that the computer use capability “has crossed the qualitative threshold,” describing it as extremely accurate, effective, and reliable.
Demonstrations showed the model’s ability to conduct deep web searches, write code, run professional 3D modeling software like Blender, and design and manage complex scenes completely autonomously on behalf of the designer.
Benchmark results and risk classification
Astra achieved unprecedented record scores on global performance benchmarks, scoring 99.9 percent on the ARC-AGI 3 test in memory alignment mode, compared to only about 7.8 percent for the GPT-5.6 model and 30.2 percent for Anthropic’s “Claude Opus 5” model. It also scored 98 percent on Level 4 of the FrontierMath test and achieved a perfect 100 percent score on the ExploitBench test, which is designed to measure the ability to develop functional security exploits from known vulnerabilities.
This latter result led OpenAI to classify Astra as the first model to reach the “critical” cybersecurity risk level, necessitating the restriction and delayed deployment of some of its more sensitive functions to ensure they are not exploited in malicious digital attacks.
Staged deployment and pricing policy
Staged deployment began on September 3 for large enterprises, then expanded to professional tier subscribers and teams via ChatGPT Work and the Codex platform. By the weekend, the model began appearing for Plus tier subscribers, with staggered access as users reported seeing it in the work interface before the public interface. The company confirmed via the X platform that it may take a few days to complete the rollout to all users, affirming that there are no current plans to make it available to free tier users.
The model is also available via the company’s API, Azure, and Amazon Bedrock, priced at $10 per million input tokens and $50 per million output tokens, which is about 2.5 times the cost of the previous model.
Mixed reactions and creator community criticism
The launch sparked a wave of enthusiasm coupled with sharp criticism. While analyses from The Chosun Ilbo pointed out that the results prove memory retention is the primary driver of performance rather than model size alone, design software users—such as those of Blender—expressed dissatisfaction with a promotional video showing a user delegating game development entirely to Astra. Even the artist responsible for the splash screen of the program’s latest release expressed disgust at the use of her work.
Early users also noted that despite high benchmark performance, the model still suffers from occasional hallucinations and instruction-following failures, raising questions about whether describing it as close to artificial general intelligence (AGI) is premature.
Frequently asked questions
Q: Who are the eligible users for the Astra model currently?
A: Subscribers to paid plans, including the $20 Plus package and professional business plans.
Q: Will the Astra model be available to users on the free tier?
A: OpenAI has confirmed there are currently no announced plans to provide the model to free tier users.
Q: What are the pricing rates for using the Astra model via cloud APIs?
A: The price is $10 per million input tokens and $50 per million output tokens via cloud platforms.
