GPT-6 Astra for business: what OpenAI's new model means for the way we work

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's flagship model for complex reasoning, coding, computer use, research and document creation. OpenAI describes it as a model for difficult end-to-end work and says it can carry out multi-step workflows across code, browsers and professional software.

OpenAI announced GPT-6 Astra on 3 September 2026. The release is staged. Enterprises in the Trusted Access Program received the first access, and OpenAI says API, Plus, Pro, Business and Enterprise access will arrive in the coming days. If Astra does not appear in your model picker, that can simply mean the rollout has not reached your workspace.

The published API specification includes a 1.05 million-token context window and up to 128,000 output tokens. Those limits matter for large source packs and long assignments, although context capacity does not remove the need for current, relevant and well-organised inputs.

What is different about Astra?

Astra combines strong reasoning with tools that let it search, inspect files, write and run code, edit files, and use a computer interface. In a suitable product or application, it can move through several parts of an assignment instead of returning one response and waiting for the next prompt.

OpenAI's model guidance says Astra is designed for multi-step workflows across browsers, code and professional software. It can continue reasoning while an external tool runs, accept updated instructions during a task and preserve more of the working context across a long piece of work. Some of these are API functions that developers must build into an application. They should not be read as a promise that every ChatGPT interface has every function on day one.

The computer-use capability is especially relevant to business. It allows a model, when given the right tool and permission, to operate a browser or software interface much as a person would: opening pages, entering information, selecting controls and checking what happened. Coding tools let it create or modify software, run tests and correct problems. Web search, file search and large context support research and document-heavy assignments.

These capabilities also increase the need for control. A model that can take several actions can repeat an error across several systems. Access should follow the person's existing permissions, consequential actions need approval, and the workflow needs a clear way to stop, recover and escalate.

From answering questions to performing work

Most business adoption began with individual tasks. A person asked ChatGPT to draft an email, summarise a report, compare documents or prepare some content. The person gathered the information, moved it between systems and decided what happened next.

Astra extends the size of the unit of work that can be delegated. The assignment might begin with a business outcome, then require research, source collection, analysis, document preparation, system updates and a final check. The model can attempt more of that sequence when the tools, instructions and permissions are available.

The practical progression is simple: AI answers questions, AI creates work, AI performs work. Each step changes the human role. When AI creates a draft, a person still coordinates the surrounding process. When AI performs several steps, people spend more time defining the outcome, providing authority, reviewing judgement calls and handling exceptions.

This does not establish a productivity gain by itself. A longer automated sequence can save effort when the inputs and rules are sound. It can also create hidden rework, weak decisions or records that are difficult to audit. The useful measure is the finished business result, including quality, elapsed time, rework and risk.

What could Astra do inside a business?

In business development, an agent could research an organisation from approved sources, identify relevant people, prepare an account brief, draft outreach and propose a CRM update. Today, models can perform many of those individual tasks and use browser or API tools in a controlled environment. A business still needs rules for source quality, personal information, CRM permissions and approval before communication is sent.

In operations, Astra could collect information from several systems, analyse exceptions, update a spreadsheet and prepare a management report. This is most credible when the workflow has stable inputs, known calculations and a named reviewer. A person should retain control where the information affects payments, employment, safety or formal commitments.

In professional services, it could review a project document set, compare it with requirements, conduct supporting research and prepare a first-pass deliverable with source references. The qualified professional remains responsible for interpreting the evidence, resolving ambiguity and approving advice.

In administration, computer use could support data entry, record reconciliation, file organisation and correspondence across browser applications. The first trials should avoid irreversible actions and use checkpoints before records are submitted, deleted or sent outside the organisation.

For software and internal tools, Astra can help build or change small applications, calculators and automations, then run tests and inspect the result. OpenAI has documented coding and computer-use support today. Production use still requires software review, security checks, testing and ownership after release.

These are workflow patterns, not claims that a standard ChatGPT Business user can run each sequence immediately. Actual behaviour depends on product access, connected tools, workspace settings, source-system permissions and the way the workflow has been built.

Which workflows can we redesign now?

The useful management question is: which workflows can we redesign because AI can perform more of the work? Model access is only one input into that decision.

Start with the workflow as it operates today. Record the trigger, inputs, decisions, systems, handovers, output and accountable owner. Identify where people apply judgement, where they mainly move information and where delays or rework occur. This exposes the parts that may suit assistance or execution.

Set an autonomy boundary for each step. The AI might prepare a recommendation, complete a reversible action, or continue until it reaches a decision that requires human authority. Define the sources it may use, when context must be refreshed and the evidence it must retain with its output.

Review capacity matters. If one manager becomes the approval point for dozens of agent actions, the new method can move the bottleneck without improving the work. Reviewers need time, subject knowledge and authority to correct the output. They also need an escalation route and a way to recover when the system changes the wrong record or loses context.

The role and team effects deserve attention too. When AI takes over coordination or preparation, people may lose visibility into decisions and colleagues may stop exchanging the informal knowledge that once travelled with the work. Workflow design needs to preserve the conversations, learning and accountability that the organisation still needs.

What Astra means for ChatGPT Business

OpenAI says GPT-6 Astra access for ChatGPT Business is coming in the days following the initial 3 September release. That wording describes a rollout rather than universal availability. Business workspace owners should check the model picker and workspace settings before planning a pilot around Astra.

ChatGPT Business gives teams a shared workspace with central administration. OpenAI states that Business workspace data is excluded from model training by default. The plan can provide a practical setting for team use, although each organisation remains responsible for information rules, connected-app permissions, human review and the suitability of each task.

For organisations already using ChatGPT, Astra creates a reason to revisit earlier use cases. A task that previously stopped at a draft may now support a longer sequence. Test the new sequence against the existing method and keep the same standard for evidence, review and accountability.

Licence activity, message counts and model use show adoption activity. Business value needs evidence from the work itself: accepted outputs, reduced elapsed time, lower rework, better service, fewer missed steps or another result the workflow owner already cares about.

What should businesses do now?

Choose a few repetitive digital workflows with clear outputs. Break each workflow into tasks and mark the points that require human judgement, approval or professional responsibility. Decide which tasks the AI may assist with and which reversible actions it may perform.

Run controlled trials with approved information, limited permissions and named reviewers. Record the source material, the actions taken and any exceptions. Compare the new method with the current one using time, output quality, rework and risk.

Use what you learn to change the workflow, permissions and review points. Keep successful methods as owned business procedures with a version and review date. Stop experiments that create more checking than benefit.

• Identify repetitive digital workflows.
• Break each workflow into tasks and decisions.
• Mark human judgement and approval points.
• Set AI assistance and execution boundaries.
• Select a small group of controlled trials.
• Measure finished work, quality, rework and risk.
• Revise the workflow from observed results.

A practical response to a more capable model

Astra expands the work that an AI system can attempt. The best response is a disciplined look at real workflows, with enough permission for useful action and enough human authority to protect the result.

Exploring what the latest ChatGPT capabilities could mean for your organisation? Addaptive helps Australian businesses move from individual AI use into adoption, workflow redesign and practical implementation.

Frequently asked questions