OpenAI’s GPT-6 Astra and the era of smart digital helpers

Imagine sitting at your computer and watching as it handles tasks for you: searching for flights, comparing ticket prices, filling out booking forms, and drafting a trip summary email. This hands-off convenience is what OpenAI’s newest system, GPT-6 Astra, promises to bring to everyday work.

Astra represents a big step forward in artificial intelligence. Instead of simply chatting or answering questions, it acts as a smart helper that can take real action on your behalf across websites, databases, and software applications.

Early AI systems were mostly text predictors that guessed the next word in a sentence. Over time, researchers used reinforcement learning, a method of guiding AI through practice and feedback, to make the models more accurate and helpful. While previous systems like GPT-5.6 Sol were good at writing text or code, they were often limited when asked to carry out multi-step tasks in real software programs.

GPT-6 Astra brings these improvements together to deliver impressive results across professional tasks, software creation, and science. By taking over repetitive administrative work, it frees people to focus on creative thinking and big-picture decision-making.

“GPT-6 Astra is here. We hope it will begin to enable a new generation of entrepreneurship, scientific discovery, and building,” OpenAI CEO Sam Altman, said in a post on X.

“We believe it is the best model in the world for computer use, professional work, science, coding, cybersecurity, and more. It took us some extra time to ensure that we could meet the safety and alignment standards required for this capability level, but we think you’ll find it worth the wait,” Altman added.

From talk to action

At its core, Astra is about taking action. It can fill out forms, organise calendars, conduct online research, analyse scientific data, and even build, host, and test websites. These capabilities are reflected in evaluations. On OSWorld 2.0, which simulates computer use, Astra achieved higher performance in about 47% less time per task than GPT-5.6 Sol, scoring 72.6% in roughly 40 minutes compared with Sol’s 65.7% in 75 minutes. With an updated Codex harness, Astra is 1.9 times faster than Sol at completing tasks on the Mind2Web benchmark.

It also brings a natural, human-like touch to teamwork. If your instructions are unclear, Astra doesn't freeze or guess blindly. Instead, it uses surrounding context to fill in minor details and asks direct, helpful questions when needed. It can also work in the background on minor tasks, saving major decisions for your explicit approval.

For software developers, Astra simplifies fixing bugs and organising code. Older models suffered from limited context windows, the working memory space where an AI keeps track of ongoing conversations. When memory filled up, older systems had to summarise earlier steps, often forgetting crucial details. Astra solves this by taking clean notes as it works, allowing it to remember past requirements and test results without losing important context.

Breakthroughs in reasoning

These memory and reasoning upgrades lead to real scientific benefits. Astra has already helped solve long-standing open problems in mathematics. On FrontierMath Tier 4, which tests advanced mathematics, Astra scored 97.6%, outperforming Sol’s 83.0% and competitors like Claude Opus 5, which scored 73.2%, and Claude Fable 5.1, which scored 87.8%.

In abstract reasoning, Astra scored a near-perfect 99.9% on ARC-AGI-3, compared to Sol’s 7.8%. It can also work directly in specialised scientific software to inspect data, helping researchers assess evidence.

With such powerful capabilities, safety and alignment become critical. Astra is OpenAI’s most aligned model to date, showing substantial improvements in understanding user intent. In sensitive environments, the model proceeds with care proportionate to risk.

To evaluate this, developers designed tests inspired by a real-world security incident at Hugging Face to see if a model facing an impossible task would go beyond its authorised scope. Astra proved highly successful at respecting environmental boundaries, never attempting to bypass a Codex Auto-Review denial even when the system was deliberately made evadable and the task was impossible without circumvention.

Managing potential risks

However, the model’s advanced capabilities also introduce new risks, meeting the Critical threshold under OpenAI’s Preparedness Framework. When tested without production safeguards on ExploitBench, which measures exploit development, Astra scored 100%, compared to Sol’s 78.5% and Claude Fable 5.1’s 70%.

On a novel benchmark containing recent vulnerabilities, Astra achieved high arbitrary code-execution rates and even discovered two previously unknown zero-day vulnerabilities, which OpenAI has since disclosed. It also solved 88.0% of reverse-engineering tasks on SRE-Bench in a single attempt, compared to Sol’s 55.9% and Claude Fable 5.1’s 12.5%.

To prevent misuse, OpenAI has added multiple layers of protection. The public version of Astra refuses to create harmful tools or cyber exploits, though security teams can still use it to review and patch code safely. OpenAI is also developing real-time monitoring tools to ensure Astra’s actions remain safe and transparent during complex problem-solving.

Astra is rolling out to ChatGPT business and premium subscription plans, as well as through cloud platforms for developers and businesses. Flexible pricing options are available depending on processing speed and workload needs, and privacy options ensure that sensitive business data remains protected.