OpenAI has started rolling out GPT-6 Astra, a new model aimed at work that runs across reasoning, software, browsers, and business documents rather than stopping at a chat answer. The first access is limited to selected organizations, with wider availability planned for ChatGPT Plus, Pro, Business, and Enterprise users and through the API, Microsoft Azure, and AWS Bedrock (OpenAI).
That rollout matters because Astra changes two inputs operators usually keep separate: capability and operating cost. OpenAI lists standard API pricing at $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million and cache writes at $12.50 per million (OpenAI API pricing). This is a model to test against a specific workflow, not a reason to migrate every prompt on launch day.
What happened
OpenAI describes Astra as state of the art for computer use, browsing, software engineering, cybersecurity, science, and professional work. Its examples include filling out forms, updating CRM records, organizing calendars, researching online, drafting summaries, creating websites, and running front-end quality checks (OpenAI).
The company reports a 72.6% score on its OSWorld 2.0 computer-use evaluation, completing tasks in about 40 minutes on average. GPT-5.6 Sol scored 65.7% at roughly 75 minutes per task in the same comparison, according to OpenAI’s launch material (OpenAI). Those are vendor-reported results, so treat them as a reason to run your own test set rather than as a production guarantee.
The model also brings a meaningful security constraint. OpenAI says Astra reaches its Critical threshold for cybersecurity under the company’s Preparedness Framework. The launch includes additional monitoring and restrictions around high-risk cyber work, while defensive uses such as secure code review and patching remain part of the intended use (OpenAI Deployment Safety Hub).
Why it matters for operators
Astra’s headline feature is not simply a higher benchmark score. It is the ability to carry a task through several systems while preserving instructions and handling changes. That puts pressure on workflows built around a sequence of small automations, especially where a human currently copies data between a browser, a spreadsheet, a CRM, and an email draft.
The cost profile still argues for restraint. At $10 input and $50 output per million tokens, an agent that repeatedly reads long records or produces verbose artifacts can become expensive faster than a short chat prompt. Requests above 272,000 input tokens are priced at higher long-context rates, according to OpenAI’s developer documentation (OpenAI API pricing). A workflow that looks cheap in a demo can have very different unit economics once it runs across a full customer queue.
There is also a control-plane question. Astra can act on a user’s behalf, but OpenAI’s own materials describe monitoring that can pause or stop a conversation when a potential safety issue is detected (OpenAI release notes). That may be appropriate for risky actions, but it means operators need clear retry behavior, human escalation, and audit logging before they give an agent write access to production systems.
What operators should do
- Choose one bounded workflow. Start with a task that has a clear input, a measurable output, and a reversible failure mode. Updating a draft CRM field or preparing a research brief is easier to supervise than sending payments or changing customer permissions.
- Build a cost baseline before switching models. Record input tokens, output tokens, tool calls, human review time, and failure retries for the current workflow. Then run the same cases through Astra. Compare cost per completed task, not cost per token alone.
- Keep actions behind approvals. Use read-only credentials for the first test. Require a person to approve external messages, financial actions, account changes, and destructive edits. Log the model’s proposed action and the final human decision.
- Test prompt injection and stale instructions. Browser and CRM agents encounter untrusted text. Put adversarial content in webpages, emails, and records, then check whether the model follows the task boundary or treats the content as a new instruction.
- Keep a fallback model. Maintain an existing model for routine work until Astra has passed your own reliability, latency, and cost checks. A staged rollout is easier to reverse than a rushed migration.
The practical next step is a two-week pilot with a fixed task set and a hard spend cap. GPT-6 Astra may reduce handoffs for complex work, but the value will come from a controlled workflow with clear permissions, not from turning on autonomy everywhere.
Related links
- Learn: /learn/how-to-build-your-ai-stack-2026
- Learn: /learn/ai-stack-for-solopreneurs-2026
- Tool: /tools/ai-model-comparison
Primary source: OpenAI’s GPT-6 Astra announcement.