OpenAI’s GPT-6 Astra is reaching $20 ChatGPT Plus users. The model first launched on September 3 for a limited group of organizations and higher-tier plans. OpenAI calls it the most intelligent and aligned model in the world. It marks a full version jump from GPT-5.6 Sol and pushes further into AI-driven computer use.
Computer Use as the Core Capability
Astra’s main feature is computer use. It can navigate software, browse the web, and run programs for the user without an API. OpenAI cofounder Greg Brockman said in a press briefing that computer use has crossed the qualitative threshold. He described it as very accurate, efficient, and reliable.
Demonstrations show the model researching topics online, writing scripts, and running 3D modeling software like Blender on its own.
Benchmark Results
On ARC-AGI-3 in its memory-retaining Adapter Harness mode, Astra scored 99.9 percent. GPT-5.6 reached 7.8 percent on the same test. Anthropic’s Claude Opus 5 scored 30.2 percent.
Astra also posted 98 percent on FrontierMath Tier 4 and 100 percent on ExploitBench. ExploitBench checks whether a model can build working exploits from known vulnerabilities. That result led OpenAI to rate Astra as its first model at the critical level for cybersecurity risk. The company responded by gating the most sensitive capabilities during rollout.
Staged Access and Pricing
The rollout started September 3 with selected organizations. It then moved to Pro, Enterprise, and Business Premium users through ChatGPT Work and Codex. As of this weekend, Astra is showing up for Plus subscribers. Access is still uneven. Some users see it in ChatGPT Work before regular Chat. OpenAI posted on X that it might take a few days to reach Plus and Business users. Free-tier users have no announced access.
The model is also available through the OpenAI API, Microsoft Azure, and Amazon Bedrock. Pricing sits at $10 per million input tokens and $50 per million output tokens. That is roughly 2.5 times the cost of GPT-5.6 Sol.
Pushback and Remaining Limits
The launch has drawn mixed reactions. Reporting from Chosun Ilbo pointed to ARC Prize benchmark results showing that memory retention, not just model size, drives performance. The finding carries implications for the semiconductor and AI infrastructure industries.
Traditional Blender users and other creatives criticized OpenAI’s marketing video. The ad showed a solitary user handing off 3D modeling and game development entirely to Astra. The artist behind Blender 5.2’s splash screen said she felt disgust at seeing her work used in the ad.
Early testers report that Astra still hallucinates and can ignore instructions. Those observations have led some to question whether framing the model as close to AGI is premature.