When ChatGPT appeared publicly in November 2022, it felt revolutionary simply because people could have a surprisingly natural conversation with a machine. The original system, based on GPT-3.5, could answer questions, explain concepts, write text, and help with code.
Less than four years later, AI has moved far beyond that model.
GPT-6 Astra can reason through difficult problems, browse information, operate computers, write and debug software, create professional documents, work across applications, and perform long multi-step tasks with much less human supervision.
That raises an obvious question: are we moving toward something resembling Skynet — a powerful autonomous intelligence capable of acting independently in the real world?
The answer is more complicated than either “yes” or “no.”
The First ChatGPT Was Primarily a Conversational Machine
The original ChatGPT launched on November 30, 2022.
Its breakthrough was not that it possessed superhuman intelligence. Its breakthrough was accessibility.
For the first time, hundreds of millions of people could interact with a capable language model simply by typing ordinary sentences.
ChatGPT could write essays, summarize information, explain mathematics, generate code, and maintain a conversation.
But it was fundamentally reactive.
You asked a question, and it produced an answer.
It did not independently navigate your computer, carry a complex project through multiple applications, or autonomously pursue objectives over long periods.
AI Learned to See, Hear, Reason, and Use Tools
The following generations expanded rapidly.
Multimodal AI gained the ability to understand images, audio, and increasingly complex documents. Reasoning-oriented models became much stronger at mathematics, science, coding, and multi-step analytical problems.
AI also gained access to tools.
This changed everything.
A language model that only generates text is limited to describing what should happen. A model connected to browsers, applications, code environments, files, databases, and computers can actually take actions.
By 2025, ChatGPT had already evolved far beyond the original research preview, with GPT-4, GPT-4o, reasoning models such as o1, and later GPT-5 representing successive major stages.
Astra Represents a Different Kind of AI
GPT-6 Astra, released in September 2026, pushes this transition much further.
Astra is designed for end-to-end work rather than isolated answers.
It can combine:
- Complex reasoning
- Computer use
- Web research
- Software engineering
- Document creation
- Data analysis
- Professional workflows
- Long-context processing
- Multi-step planning
Its documented context window exceeds one million tokens, allowing the model to process extremely large collections of information within a single working context.
This represents a profound change.
The unit of AI assistance is shifting from “answer this question” toward “complete this objective.”
The Biggest Leap Is Agency
Raw intelligence is only part of the story.
The more important development may be agency.
An agentic AI can decide which intermediate steps are necessary to accomplish a goal.
Imagine asking an early ChatGPT:
“Analyze these companies and prepare an investment presentation.”
You might have needed to manually provide information, request analysis, ask for separate sections, create charts yourself, and transfer everything into presentation software.
An advanced agent can potentially research the companies, compare financial information, organize findings, create tables, prepare slides, inspect the result, and revise it.
That is much closer to the way a human employee completes a task.
Why Astra Raises the Skynet Question
In the Terminator fictional universe, Skynet is not merely an intelligent chatbot.
It combines several abilities:
autonomous decision-making, strategic planning, access to critical infrastructure, military control, large-scale computer connectivity, and the ability to pursue objectives without meaningful human supervision.
Modern AI has now developed pieces of that technological puzzle.
Advanced models can reason.
They can write software.
They can operate computers.
They can coordinate multi-step actions.
They can search networks.
And increasingly, they can interact with physical machines through external systems.
That makes comparisons with science fiction understandable.
But having some components associated with fictional autonomous AI is not the same as having Skynet.
The Cybersecurity Milestone Is Especially Important
Astra illustrates how serious AI capabilities have become.
OpenAI classifies GPT-6 Astra as its first model reaching the Critical cybersecurity capability level under its Preparedness Framework.
With appropriate tools and access, the model can identify previously unknown security vulnerabilities and develop exploitation strategies across protected systems without requiring a person to guide every individual step.
That is a significant milestone.
A few years ago, mainstream generative AI mainly helped programmers write code.
Now frontier systems can perform sophisticated cybersecurity reasoning and autonomous technical workflows.
This does not make Astra Skynet, but it demonstrates why control, permissions, monitoring, and security become increasingly important as AI capabilities improve.
What Current AI Still Lacks
A Skynet-like system would require more than intelligence.
It would need persistent autonomous goals.
It would need reliable long-term planning across changing environments.
It would need broad and largely unrestricted access to infrastructure.
It would need the ability to obtain resources, maintain itself, replicate, survive attempts to shut it down, and continuously pursue objectives without external authorization.
Current commercial AI systems are not deployed that way.
Their actions depend on tools, permissions, computing infrastructure, software environments, and access granted by humans or organizations.
Even extremely capable models remain capable of errors, misunderstanding instructions, or making poor judgments.
Intelligence and autonomy are separate variables.
A brilliant system without access cannot control much. A less intelligent automated system with enormous access can still create serious risks.
How Close Are We to AGI?
AGI — artificial general intelligence — usually refers to AI capable of performing a very broad range of intellectual tasks at or above human level.
There is no universally accepted test that determines exactly when AGI has arrived.
Astra complicates the debate because its abilities already overlap with professional human performance across many domains.
On ARC-AGI-3 evaluations, the ARC Prize Foundation reported Astra reaching roughly human-level action efficiency across most evaluated levels.
That does not prove that Astra possesses human-like general intelligence.
Benchmarks measure particular abilities, while humans operate in an enormously richer physical, social, emotional, and cultural environment.
Nevertheless, the distance between specialized machine intelligence and broadly useful machine intelligence has clearly narrowed dramatically.
Expert Perspective
Greg Kamradt of the ARC Prize Foundation described Astra’s performance as a meaningful step change in frontier AI, particularly in its ability to navigate unfamiliar environments and solve new problems efficiently.
This highlights one of the most important developments in modern AI.
The frontier is no longer simply about memorizing more information.
Researchers increasingly care about whether an AI can adapt, reason, plan, use tools, recognize mistakes, and complete unfamiliar tasks.
Those capabilities are much closer to what people traditionally associate with general intelligence.
The Real Risk May Look Less Like Science Fiction
A future dangerous AI does not necessarily need killer robots or dramatic self-awareness.
More realistic risks could involve automated cyberattacks, manipulation at enormous scale, poorly supervised autonomous systems, financial disruption, military decision-support errors, or AI agents receiving excessive access to important infrastructure.
The critical issue is therefore not whether an AI suddenly “wakes up.”
It is whether powerful systems are given capability, autonomy, access, and poorly defined objectives at the same time.
This is why modern AI safety increasingly focuses on monitoring, alignment, access control, sandboxing, authorization, and human oversight.
How Far Have We Really Come?
Between late 2022 and 2026, AI moved from impressive conversational software toward systems that can perform substantial portions of professional digital work.
That is an enormous leap for less than four years.
The first ChatGPT mainly communicated.
Astra can increasingly reason, investigate, create, operate software, and execute workflows.
We are still far from the complete fictional package represented by Skynet. Yet several individual capabilities that once belonged purely to science fiction are already appearing in real systems.
The important question may therefore be shifting from:
“Can machines become powerful enough?”
to:
“How much authority should increasingly capable machines be allowed to exercise?”
Interesting Facts
- ChatGPT was publicly launched on November 30, 2022.
- The original public ChatGPT used GPT-3.5.
- GPT-6 Astra was released in September 2026.
- Astra supports a context window of about 1.05 million tokens through the API.
- Astra can operate browsers and computer interfaces as part of multi-step workflows.
- It is OpenAI’s first model classified at the Critical level for cybersecurity capability.
- Modern AI agents can combine reasoning with tools instead of merely generating text.
- Progress from early ChatGPT to Astra occurred in less than four years.
- High intelligence does not automatically imply unrestricted autonomy.
- A system’s real-world power depends heavily on what tools, permissions, networks, and machines it is allowed to access.
Glossary
- Artificial Intelligence (AI) — Computer systems capable of performing tasks associated with human intelligence.
- AGI — Artificial General Intelligence, a proposed form of AI capable of performing a broad range of intellectual tasks at approximately human level or beyond.
- GPT — Generative Pre-trained Transformer, a family of language and multimodal AI models.
- GPT-6 Astra — An advanced OpenAI model designed for complex reasoning, computer use, research, coding, and long multi-step workflows.
- AI Agent — An AI system capable of planning and performing multiple actions toward a goal.
- Agency — The ability of a system to select and perform actions in pursuit of an objective.
- Multimodal AI — Artificial intelligence capable of processing several types of information, such as text, images, audio, and video.
- Context Window — The amount of information an AI model can process within one working context.
- AI Alignment — The effort to make AI behavior reliably follow intended human goals and constraints.
- Autonomy — The degree to which a system can operate without direct human instruction.
- Cybersecurity Capability — An AI system’s ability to analyze, detect, exploit, or defend against computer-security vulnerabilities.
- Skynet — A fictional autonomous military artificial intelligence from the Terminator franchise, commonly used as a cultural metaphor for uncontrolled machine intelligence.
