Whenever a new frontier model arrives, the practical enterprise question is not simply what the model can do. It is what changes in real work, and what business value can be created from that change. GPT-6 Astra became generally available in Microsoft Foundry on 3 September 2026, with availability spanning across Copilot Cowork (Microsoft Copilot), Copilot Studio and GitHub Copilot.

For Infosys Topaz, the significance of GPT-6 Astra lies in how its advances translate into enterprise value. With its composable, open and interoperable architecture, Infosys Topaz Fabric enables newer models to be incorporated without rebuilding the underlying AI stack. The opportunity, therefore, is to identify where GPT-6 Astra can deepen automation, improve execution and unlock greater value through Infosys Topaz.

What Actually Changes?

GPT-5.6 Sol already supported reasoning, structured outputs, text and image inputs, function and tool calling, computer use, multi-agent orchestration in preview, and a 1.05M-token context window. GPT-6 Astra is therefore not a new agent architecture or a new platform paradigm. It is a higher-performing frontier model added to an existing stack, and the capabilities that stand out are computer use, sustained agent execution, complex engineering work and long-context effectiveness.

OpenAI’s published evaluations show computer use improving from 65.7% to 72.6% on OSWorld 2.0 and from 76.9% to 92.7% on ScreenSpot-Pro; sustained agent execution rising from 18.1% to 41.4% on AutomationBench; and complex engineering work moving from 37.3% to 57.9% on Terminal-Bench 4.0, with reported database migration tasks improving from 42.7% to 63.9%. Long-context effectiveness rises from 73.8% to 96.3%, while the context window itself remains 1.05M tokens. That distinction matters: what has improved is the ability to use information deep within the window, not the amount that can be held in it.

Taken together, these gains point to a larger shift. AI is becoming better at completing longer and more complex pieces of work with less human intervention, whether that work happens in a terminal, through a software interface or across a large body of information.

From Assistance to Execution

Enterprise AI can be viewed across three levels: assistive AI, supervised execution and delegated execution. In assistive AI, the model drafts or recommends while a person performs the work. In supervised execution, the agent acts while a person approves consequential steps. In delegated execution, the agent completes a defined task, and a person reviews the outcome.

The value of GPT-6 Astra is its potential to move suitable workloads further along this continuum. For Infosys Topaz, the opportunity is tangible:

• Computer use: supporting agentic legacy modernization, including systems without dedicated APIs.

• Complex coding and terminal work: strengthening SDLC agents, testing and application management.

• Long-context effectiveness: improving discovery and modernization assessment across large, undocumented estates.

• Sustained multi-step execution: reducing human handoffs across Infosys Topaz Fabric agent flows.

• Artifact quality: supporting Process AI and back-office work, shifting effort from extensive rewriting toward expert review.

The potential business value is greater productivity, faster software and process transformation, deeper agentic automation, improved output quality and better economics per completed business task.

Measure Completed Work, Not Just Tokens

A more capable model does not automatically mean better economics. The relevant measure is cost per completed business task, not simply cost per million tokens. GPT-6 Astra can complete more of what it attempts, but the published figures cited here use maximum reasoning effort, which consumes more tokens and time per attempt. Economics improve only when higher completion rates, better quality and reduced human intervention outweigh the additional execution cost.

This is also why evaluation matters as much as capability. Benchmark results are not uniform across task types, and strong performance on one evaluation does not reliably predict performance on another. The model’s classification at OpenAI’s Critical cybersecurity capability threshold is equally material to how security-related agent work is scoped. Published benchmarks tell us where to look. Evaluation of a real client workload is what should decide the architecture.

Greater Capability Demands Greater AI Trust

Greater autonomy also makes governance and containment more important. With computer use, information presented through an interface can be incomplete, misleading or deliberately designed to influence an agent, which makes scoped credentials, approved resources, human checkpoints for consequential actions and proper activity records essential. Microsoft Foundry provides controls across identity and access management, encryption, private networking, role-based access, content filtering, safety evaluation and monitoring, and prompts and outputs are not used to train the models. Applying these controls within the context of a client’s data, applications and processes is where AI Trust becomes critical, aligning with the Infosys AI-first value framework. As agents become more capable of acting, the boundaries within which they operate must become equally deliberate.

Alongside trusted execution, enterprises need the flexibility to adopt new AI capabilities as they emerge. Infosys Topaz enables this through an adaptable AI foundation, with Infosys Topaz Fabric providing the composable, open and interoperable architecture to integrate emerging models and capabilities. The Agentic AI Foundry and Infosys’ collaboration with OpenAI further strengthen this foundation, enabling advances such as GPT-6 Astra to support deeper agentic automation and enterprise transformation.

Conclusion

GPT-6 Astra represents a meaningful upgrade in the areas that matter to agentic execution. But enduring enterprise value does not sit in the model alone. It sits in the client’s data, processes and applications, the governance surrounding AI, and the judgement required to know when an outcome is right or wrong. Infosys Topaz is where these elements come together, helping translate advances in frontier intelligence into trusted and measurable enterprise value. Frontier models will continue to change. The ability to put them to work responsibly, effectively and economically is what lasts.