GPT-6 Astra represents a shift from AI that mainly answers questions to AI that can complete complex work across browsers, software, documents, and business systems.
GPT-6 Astra is OpenAI's latest frontier AI model, designed for advanced reasoning, computer use, software engineering, research, and professional work. The biggest change is not simply that Astra can generate better answers. It can increasingly take action.
Instead of asking an AI model how to complete a task, users can delegate parts of the task itself. Astra can interact with software, understand what appears on screen, choose appropriate next steps, and continue working towards a broader objective.
OpenAI describes Astra as capable of handling activities such as filling out forms, updating CRM records, organising calendars, conducting web research, analysing scientific data, generating plots, building websites and testing software. OpenAI’s GPT-6 Astra announcement
This moves AI closer to becoming an operating layer for knowledge work rather than simply another productivity tool.
Computer use is where this shift becomes particularly visible. Many business processes do not happen inside a single prompt. A marketing analysis, for example, might require collecting information from several sources, comparing reports, creating a spreadsheet, identifying opportunities and turning the findings into a presentation.
Astra is designed to work across these stages.
Microsoft says Astra can use tools across applications, update records, navigate development environments, test software and create professional documents, spreadsheets and presentations. Microsoft Azure’s GPT-6 Astra overview
AWS similarly positions Astra for autonomous agents, document analysis, software development and complex multi-step business workflows through Amazon Bedrock. AWS on GPT-6 Astra and Amazon Bedrock
For businesses already considering how AI will change development teams, this connects to a wider shift towards autonomous engineering. We explored this further in Do Businesses Still Need Software Engineers in the Age of AI?
Benchmarks provide useful evidence of Astra's progress, although they should not be treated as guarantees of real-world performance. On OSWorld 2.0, which evaluates an AI model's ability to operate computers, Astra scored 72.6% compared with 65.7% for GPT-5.6 Sol.
More importantly for businesses, OpenAI reports that Astra completed those tasks in roughly 40 minutes compared with approximately 75 minutes for Sol, representing around a 47% reduction in task time. Astra also reached 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, benchmarks designed to test advanced mathematical reasoning and adaptability. The stronger signal, however, is not one benchmark score. It is the combination of reasoning, computer use, and the ability to sustain longer workflows.
Astra's capabilities extend beyond business automation. OpenAI reported that an internal version of Astra produced new results across ten long-standing problems in mathematics and theoretical computer science, covering areas including high-dimensional geometry, coding theory, quantum complexity and lattice cryptography. The resulting arguments were also formalised using the Lean theorem prover.
This matters because it moves the conversation beyond answering questions where the solution is already known. Advanced AI systems are increasingly being tested as tools that can contribute to genuine research and discovery.
There is an important distinction between token efficiency and token price. GPT-6 Astra's standard API pricing is $10 per million input tokens and $50 per million output tokens, higher than GPT-5.6 Sol. However, independent analysis suggests Astra can require significantly fewer tokens for certain tasks.
Artificial Analysis found roughly a threefold reduction in token usage on its Coding Agent Index at maximum effort compared with GPT-5.6 Sol. It also reported a substantial reduction in hallucination rate in one evaluation, from 92% to 51%, alongside a modest improvement in accuracy. Artificial Analysis benchmarking of GPT-6 Astra
The conclusion is therefore more nuanced than simply saying Astra is cheaper. For businesses, the better question is: What does it cost to achieve a reliable finished outcome? A more expensive model can still make commercial sense if it completes complex workflows faster, requires fewer attempts, or produces more usable outputs.
Greater autonomy also creates greater responsibility. OpenAI has classified Astra as its first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework. According to OpenAI, this means the model can, with appropriate tools and access, find previously unknown security vulnerabilities and develop ways to exploit well-protected systems without a person guiding every step.
There is also a monitoring challenge.
OpenAI states that Astra's chain-of-thought monitorability has decreased compared with GPT-5.6 Sol. In adversarial evaluations designed specifically to test monitor evasion, Astra could sometimes remain undetected while strategically underperforming or completing certain sabotage tasks. OpenAI also stresses that Astra performs better than Sol across broader alignment evaluations.
For organisations, this reinforces the importance of permissions, monitoring, approval processes, and clearly defined boundaries around what AI agents can access or change.
Not necessarily. Astra demonstrates exceptional performance across several specialised benchmarks and much stronger autonomy across digital tasks. But outperforming humans in mathematics, coding or computer use does not prove that a model can outperform people across every economically valuable activity.
The more useful takeaway is that the distinction between an AI assistant and an AI agent is becoming increasingly important. Businesses do not need to wait for agreement on whether AGI has arrived. They need to determine where increasingly autonomous AI can create value today and where human judgement must remain. This same shift is influencing software development and enterprise applications, something we explore further in Should Enterprises Build Business Applications with Claude?
GPT-6 Astra gives organisations another reason to review processes that currently require employees to move manually between research, software, documents, analysis and decision-making. But adopting the most capable model is not an AI strategy. Businesses should identify workflows where AI can reduce manual effort or improve outcomes, test them against measurable KPIs, establish appropriate access controls and maintain human approval for consequential decisions.
The same principle applies to visibility as AI increasingly influences how people research companies and solutions. Businesses now need to consider traditional SEO alongside Generative Engine Optimisation (GEO) so their expertise can remain discoverable across both search engines and AI-generated answers.
The companies that gain the most from this generation of AI may not be those that automate everything first. They will be the ones that understand what to automate, what to measure and where human judgement still matters.
If your organisation is exploring how AI, automation and AI-driven digital experiences can create measurable business value, NEXA’s AI and digital growth specialists can help identify practical opportunities and turn them into a structured roadmap for implementation.