GPT-6 Astra moves from chat directly into business applications. OpenAI is betting on work without integrations

OpenAI is positioning GPT-6 Astra not merely as another more capable language model, but as a work layer above enterprise applications. In ChatGPT Work and Codex, the model is supposed to write code, browse the web and operate software through the user interface, including when a system offers no application programming interface. That lowers the initial technical barrier: some automation can be introduced without first rebuilding corporate systems. The first customer tests published by OpenAI show improvements mainly on long, multi-step tasks. Databricks reports better cost per completed task than with GPT-5.6 Sol. Hebbia measured 17 percent better instruction following and a 19 percent higher rate of correctly locating the source document. Databricks' Ivan Zhou summarizes the result by saying that it offers substantially better cost per task. These are, however, measurements from partners selected by the vendor rather than an independent audit. The economics of deployment will therefore not be determined by the price list alone. OpenAI set pricing from $10 per million input tokens and $50 per million output tokens. In agentic workflows, however, costs also arise from repeated steps, tool use, checks and human oversight. A company should track the cost per completed process, the number of human interventions and the share of tasks returned for correction. Only these indicators will show whether the new model actually reduces costs. The second part of the offering consists of safety features. Enterprise administrators can restrict permitted websites and applications, control file uploads and require confirmation before an important action. In its internal testing, OpenAI reports 89 percent fewer unintended outcomes than with GPT-5.6 Sol. Such a figure is useful as a signal, not a guarantee: actual risk depends on data access, permissions and the quality of the specific workflow. For company leadership, the important shift is where the return on AI will be decided. It is no longer enough to compare model answers in chat. Value is moving toward whether a system can safely complete an accounting check, make a change in an internal system or prepare a client deliverable. Procurement and security teams therefore need a common evaluation framework: performance, cost per outcome, auditability, the ability to revoke permissions quickly and a plan for cases in which an automated step fails. For pilot projects, this leads to one more practical change: the test set has to reflect real company use cases rather than general questions. For each process, the correct outcome, unacceptable errors and the speed with which an action can be reversed should be defined in advance. Without such a baseline, greater model capability can easily be mistaken for a nicer-looking output even though operational risk or actual cost has not improved. KEY TERMS Agentic AI: A system that not only answers but plans steps, uses tools and performs actions. Cost per task: The total cost required to achieve a completed result, not just the price of one model call. Permissions: Rules defining which data and applications the system is allowed to access.

GPT-6 Astra moves from chat directly into business applications. OpenAI is betting on work without integrations

The full article is available to users registered at Hard Skills