OpenAI’s Astra model and how close it might be to AGI

OpenAI has quietly revealed what may be its most important AI system yet: a new internal model family called Astra. Instead of a simple upgrade to GPT‑4, Astra looks more like the foundation for the next generation of AI agents—and possibly the basis for whatever OpenAI eventually calls GPT‑6.

What is Astra and how is it different?

OpenAI describes Astra as its “next major model family,” not just a minor iteration. While the company hasn’t confirmed whether Astra will launch as GPT‑6, GPT‑5.x, or under a completely new name, it’s already being tested internally and shown to policymakers.

According to reports, Sam Altman has been demoing Astra in Washington, D.C. to members of the Trump administration and Congress. Those demos focused on multiple AI agents working together on complex tasks—planning, coding, debugging, and reasoning over long-running projects, rather than just answering one-off questions.

Astra’s 10 breakthroughs in math and theoretical computer science

The biggest public reveal so far is a new OpenAI research blog announcing that an internal version of Astra has produced advances on 10 open problems in mathematics and theoretical computer science. These weren’t toy exercises or exam-style questions—they were longstanding research problems that had seen little or no progress for at least a decade.

The problems span several deep areas, including:

• Group theory
• Coding theory
• Sphere packing
• Circuit complexity
• Quantum complexity
• Operator algebras

One standout example is a proof that non-sofic groups exist—a famous open question in group theory that many mathematicians have wrestled with for years. OpenAI says Astra generated the key mathematical ideas, while human researchers prepared the papers and verified the results.

AI-generated proofs and a new kind of authorship

OpenAI did something notable in how it presented these results. Instead of simply publishing polished human-written proofs, the team released:

• Formal Lean proofs used to verify each result
• Detailed reasoning walkthroughs
• The model’s own narration of how it arrived at its ideas

Crucially, the blog emphasizes that the core mathematical insights came from Astra itself. Humans checked and formalized the work, but they did not originate the ideas. OpenAI even argues that claiming human authorship would misrepresent how the discoveries were actually made.

This is one of the clearest public acknowledgements from a major AI lab that an AI system is making original research contributions, not just serving as a tool to speed up human thinking. That’s a major psychological and scientific milestone.

From snake games to solving open problems in four years

The Astra announcement hits harder when you zoom out and look at how quickly capabilities have grown. Just a few years ago, getting an AI model to code a simple snake game felt magical. GPT‑4 generating a basic game from a prompt was considered a breakthrough in coding ability.

Today, small open models with only a couple billion parameters can do that in seconds. What used to feel like cutting-edge capability is now the baseline.

OpenAI notes that in 2022, earlier ChatGPT models could still make basic numerical mistakes—like incorrectly comparing four-digit numbers. Fast forward four years, and an internal successor is helping to resolve decade-old research questions in advanced math and theoretical computer science. That’s a staggering jump in reasoning ability in a very short time.

The coding curve: from seconds to days of autonomous work

A key way to understand this shift is to look at how long a coding or software engineering task an AI can complete autonomously with high reliability. One analysis (from the AI research platform Meter) tracks how models have progressed on tasks they can finish end-to-end with around an 80% success rate.

• In 2022, models could only handle tasks that took humans a few seconds.
• GPT‑4 extended that to tasks taking several minutes.
• Newer models like GPT‑4.5 and Claude can now tackle tasks equivalent to roughly 30 minutes of focused human work, with minimal human intervention.

The important part: this curve doesn’t look linear. Each generation isn’t just a bit better; it can autonomously handle much longer and more complex tasks. Meter labels future milestones as “Agent 0,” “Agent 1,” and “Agent 2”:

• Agent 0: roughly where today’s top frontier models sit—handling minutes of work.
• Agent 1: systems that can work autonomously on a project for days.
• Agent 2: systems that can push through work equivalent to months or even years of human effort.

Astra appears to be one of the first serious steps toward that Agent 1 world.

Astra as a true AI agent platform

What makes Astra feel different is not just raw intelligence, but how it behaves as an agent. In Washington demos, Sam Altman reportedly showed multiple AI agents collaborating on complex projects—planning, writing code, debugging, researching, and coordinating over long time horizons.

This is a shift from “chatbot that answers questions” to “autonomous collaborator that can own a project.” Imagine an AI that can:

• Break down a multi-day task into steps
• Write and test code over many iterations
• Look up relevant research and synthesize it
• Keep track of context and goals over hundreds or thousands of actions

That’s the kind of behavior Astra is hinting at. If you’ve been following the broader AGI race, this push toward agentic systems mirrors what other labs are doing with models like xAI’s Grok series and OpenAI’s own GPT‑5.x line. For context on how other players are thinking about AGI-scale systems, see our deep dive on Grok 4.5 and xAI’s 6-trillion-parameter AGI plan.

Why these breakthroughs matter for science and industry

Right now, Astra is only available to OpenAI researchers. But if (or when) a public or API version launches, the impact could be enormous. Imagine thousands or millions of people pointing Astra-like systems at problems that have been stuck for years:

• Drug discovery and protein design
• New materials and chemistry
• Physics and cosmology
• Computer science and cryptography
• Engineering and complex systems design

Many of these fields are bottlenecked not just by human creativity, but by time and compute. If an AI can reason at a high level and work autonomously for days, it could explore far more possibilities than any single research team ever could.

OpenAI estimates that the 10 mathematical breakthroughs Astra produced cost about $2,000 worth of compute at current API pricing—roughly the price of a high-end laptop. That’s a stunning comparison: years of human thought versus a few thousand dollars of AI runtime.

Is Astra actually GPT‑6?

OpenAI hasn’t confirmed branding yet, but many in the AI community suspect Astra will effectively become the GPT‑6 model family, or at least its core. Prediction markets like Polymarket are already assigning a significant probability that Astra (or a public version of it) will launch soon, possibly within a month.

We’ve already seen OpenAI roll out intermediate models like GPT‑5.4 Pro, which some benchmarks suggest may be the current smartest widely available model. You can read more about that in our breakdown of OpenAI’s GPT‑5.4 Pro and how it compares to other frontier models. Astra looks like the next big leap beyond that tier.

Whether OpenAI markets it as GPT‑6 or under a new name, the technical story is the same: a model family that can not only chat and code, but also generate new scientific knowledge and sustain long, agentic workflows.

How close are we to AGI?

The word “singularity” gets thrown around a lot, and it’s easy to get lost in hype. But the Astra results are a concrete signal that we’re entering a new phase of AI capability.

In just a few years, we’ve moved from:

• Models that struggle with simple numerical comparisons
• To models that can code small apps
• To models that can autonomously handle 30-minute tasks
• To internal systems that help resolve decade-old research problems and may soon run multi-day projects as agents

No one can say exactly where the line to AGI is, but Astra looks like a major step toward systems that don’t just assist human experts—they actively discover, prove, and build alongside them.

If OpenAI’s internal results hold up under wider scrutiny, Astra may be remembered as one of the first models to cross that psychological threshold: from “very smart tool” to “genuine research collaborator.”

What to watch next

Over the coming months, key questions to watch include:

• How much of Astra’s capability will be exposed to the public or via API?
• Will OpenAI position it as GPT‑6, a GPT‑5.x upgrade, or a new brand entirely?
• How do independent mathematicians and computer scientists evaluate Astra’s proofs over time?
• How quickly do agent-style workflows built on Astra show up in real products and developer tools?

Whatever the branding, Astra signals that the frontier of AI is shifting from better chatbots to powerful, persistent agents that can reason deeply and work for long stretches of time. If that trend continues, the next few years of AI research and applications could look very different from anything we’ve seen so far.