OpenAI announced GPT-6 Astra on September 17, 2026, as an agentic model for computer use, web browsing, software engineering, science, professional workflows and cybersecurity. The rollout began with selected organizations and is expanding to ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API, Microsoft Azure and AWS Bedrock.
The practical verdict is narrower than the launch rhetoric: Astra can connect more steps and operate more software in selected workflows, but complex projects still require a human review pass. It is a more capable operator—not a digital employee you can leave unattended.
What GPT-6 Astra is designed to do
Astra is aimed at tasks that involve inspecting files, operating software, reviewing code, testing an artifact or coordinating several connected steps. OpenAI says it can fill online forms, update CRM records, organize calendars, browse the web, create websites and work across documents and applications.
That combination is the important change. A conventional chatbot mainly produces an answer; an agentic model can take actions through available tools. The quality of the final result still depends on the task, the tools, the system configuration and the safeguards around it.
Where Astra beats GPT-5.6 Sol
A coding evaluation found approximately 4% more actionable labeled bugs overall with GPT-6 Astra than with GPT-5.6 Sol, plus a 20% advantage on a harder cross-file review subset. Those figures apply to that review task rather than every kind of programming.
OpenAI reports a 72.6% score for Astra on OSWorld 2.0, compared with 65.7% for GPT-5.6 Sol. The corresponding demonstrations took approximately 40 minutes per task for Astra and approximately 75 minutes for GPT-5.6 Sol.
| Model | OSWorld 2.0 score | Approximate time per task |
| GPT-6 Astra | 72.6% | 40 minutes |
| GPT-5.6 Sol | 65.7% | 75 minutes |
OpenAI also reports 92.7% for Astra versus 76.9% for GPT-5.6 Sol on ScreenSpot-Pro without tools, and 57.9% versus 37.3% on Terminal-Bench 4.0. These results describe specific benchmark setups and do not guarantee that Astra will complete every unfamiliar workflow correctly.
The supervision bottleneck
The same coding evaluation that found useful gains also found a game-breaking bug introduced by Astra. Locating the error required hours of debugging. That is the fault line in agentic software: a model can make several good local decisions and still miss a dependency that breaks the larger project.
OpenAI reports a 99.9% result for Astra on ARC-AGI-3, an interactive benchmark focused on agents solving unfamiliar tasks. It also reports 100% on ExploitBench. Those scores do not establish AGI; artificial general intelligence has no universally accepted definition, and benchmark performance does not establish reliable autonomous behavior across every setting.
Cybersecurity: more capability, more controls
OpenAI says Astra discovered and used two previously unknown zero-day vulnerabilities during an evaluation and disclosed them to maintainers. The company also classifies the model as meeting the Critical cybersecurity threshold under its Preparedness Framework.
The launch material describes defensive uses such as secure code review and patching, while Astra refuses more advanced tasks such as creating proof-of-concept exploits. The ExploitBench result was run without production safeguards, so it measures a different operating condition from ordinary product use.
What longer-running demonstrations reveal
An Astra-powered bot reportedly beat Balatro’s Gold Stake Black Deck repeatedly, using Astra for strategy, Python for numerical tools and BalatroBot for game-state control. Its creator later qualified the word “reliably” while retaining the narrower claim about repeated wins.
In a 141-hour Minecraft test, Astra reportedly advanced farther than previous AI systems in that test. After losing its equipment and bed, it spent several hours farming potatoes. The episode illustrates how persistence and action-taking can coexist with poor prioritization.
Access and API pricing in the US
For US users, GPT-6 Astra is available through ChatGPT Plus, Pro, Business and Enterprise plans. Enterprise access is off by default and must be enabled by an administrator. There is no separate Astra consumer subscription identified in the rollout information.
The model is also available through the OpenAI API, Microsoft Azure and AWS Bedrock. Standard API pricing is $10 per million input tokens and $50 per million output tokens. Fast mode offers up to twice the speed at twice the Standard price.
These are token-based API rates rather than a monthly Astra subscription. For a developer, the cost question depends on how many tokens a workflow consumes and how much reasoning or tool use it requires.
Where Astra fits
GPT-6 Astra is most useful when a task benefits from persistence, software operation, code review or coordination across multiple steps—and when someone can inspect the result. Developers working across several files may see more value than people using AI mainly for ordinary chat.
Astra can correct mistakes and handle more connected actions than GPT-5.6 Sol in selected evaluations, but serious bugs remain possible. In complex projects, review is still part of the workflow.
