OpenAI introduces GPT-6 Astra for business workflows, computer use and coding

OpenAI says it has launched GPT-6 Astra, a business-focused model designed to handle computer use, software engineering, browsing and other professional tasks inside existing workplace applications. The model is available through ChatGPT Work, Codex and the OpenAI API, although enterprise access is off by default at launch.

The announcement positions Astra as a model intended to reduce the preparation typically required before companies deploy AI. According to OpenAI, it can operate applications that do not expose an API, allowing teams to use existing workflows rather than first rebuilding them around custom integrations. That promise places the release in direct competition with enterprise AI platforms focused on automation, coding and agentic software use.

What OpenAI says Astra changes

OpenAI describes GPT-6 Astra as its most capable model for business work, with improvements in reasoning, writing, design judgment and the ability to interact with software through a computer interface. In ChatGPT Work and Codex, the company says Astra can write code and navigate the same applications employees already use, including tools that lack conventional API access.

The company also says Astra is better at following organizational voice, templates and design standards. That could matter for product and marketing teams that currently spend significant time revising AI-generated presentations, documents and other branded materials before they are usable.

OpenAI cited several early examples from unnamed customers, including GPU optimization, identifying discrepancies in financial statements and producing presentations that better matched brand requirements. These examples are customer-use claims supplied by OpenAI; the announcement does not provide independent verification, customer names, deployment sizes or measured business outcomes.

The model is also available in the OpenAI API, with pricing starting at $10 per million input tokens and $50 per million output tokens, according to the company. Enterprise administrators can enable Astra under their existing rate card and agreement, while eligible API customers can receive Zero Data Retention on supported endpoints subject to approval.

Evidence behind the performance claims

The strongest performance claims in the announcement are vendor-reported. OpenAI says Astra reaches state-of-the-art results in computer use, browsing, professional work, software engineering, cybersecurity and science, but it does not publish a complete set of evaluation results in the supplied material.

OpenAI specifically cites Terminal Bench 4.0 and the Artificial Analysis Intelligence Index when describing cost efficiency on professional work and coding evaluations. It says Astra was trained to complete tasks with fewer tokens and retries, potentially reducing rework and the cost of individual tasks. Those claims should be treated as benchmarks selected and characterized by OpenAI until independent testing shows how the model performs across comparable workloads.

The company also described an internal engineering example. OpenAI says its team used Astra to identify a memory-allocation bottleneck affecting Codex sessions in a test environment. After changing memory allocators, the team reported 25-times lower turn latency while peak memory use rose by roughly 30 percent. This is an internal case study, not evidence that the same improvement will appear in customer deployments.

OpenAI says it tested Astra on an internal computer-use safety benchmark involving scenarios such as exposing confidential information, broadly sharing a dashboard or deleting data. The company reports that Astra produced unintended outcomes 89 percent less often than GPT-5.6 Sol and 74.7 percent less often than Claude Fable 5.1. Because the benchmark is internal and the supplied announcement does not describe its full methodology, the comparisons cannot be independently assessed from this release alone.

The source base for this story is also narrow. The available wire item contains no article text, while the detailed account comes from OpenAI News. As a result, adoption, reliability and competitive-performance claims remain primarily the company’s own account rather than independently reported findings.

Controls for enterprise deployment

Computer use can make an AI system more useful, but it also gives mistakes a path into business systems. OpenAI says Astra includes stronger adherence to human intent and authorization, alongside controls intended to limit what the model can access and do.

New administrator settings can restrict access to approved websites and desktop applications, manage uploads and downloads, and control browsing history. ChatGPT Work and Codex also support confirmation policies that require human approval before consequential actions. OpenAI says automated review can inspect potentially unsafe or unauthorized tool calls.

These controls support a staged rollout: an organization could begin with a narrow set of applications, limited permissions and approval requirements before expanding access. That is more practical for enterprise AI buyers than granting an agent unrestricted browser or desktop access from the start, but the announcement does not establish how these safeguards perform under real-world attack or failure conditions.

OpenAI is also introducing enterprise plugins in ChatGPT Desktop for Oracle Analytics, Power BI, Navan and Avalara. The company says the plugins use its latest browser-use capabilities to connect users with familiar enterprise applications. Their inclusion suggests OpenAI is pursuing both broad computer interaction and more controlled connections to specific business systems.

Astra is described as the first model to reach the Critical cybersecurity capability threshold in OpenAI’s Preparedness Framework. OpenAI says it strengthened protections against misuse, unauthorized actions and attempts to bypass safeguards. The framework designation signals that the company considers the model capable of materially stronger cyber-related work, but it also highlights why deployment controls and monitoring will be important.

What the launch means for builders and buyers

For product teams, the most consequential part of the release is not simply a higher benchmark score. It is the claim that GPT-6 Astra can operate existing software without requiring an API integration. If reliable, that could shorten the path from an internal experiment to an automated workflow in finance, operations, marketing or engineering.

The trade-off is that computer-based interaction is less predictable than a tightly defined API call. Screen changes, permissions, ambiguous instructions and unexpected application states can create failure modes that are difficult to reproduce. Builders will need audit logs, approval gates, recovery paths and clear limits on data movement before allowing Astra to perform consequential work.

Pricing will also require workload-level analysis. Lower token use and fewer retries could reduce the cost of a task, but computer-use workflows may consume more time and involve additional review or tool calls. Enterprises should compare the full cost of completion, including human approvals, monitoring and failed actions, rather than relying only on token rates.

For software teams, Codex access and the reported internal latency result may make Astra attractive for debugging, code generation and engineering support. However, buyers should validate performance against their own repositories, security policies and deployment environments. OpenAI’s reported results do not establish that the model will improve every development workflow.

What to watch next

The next useful signals will come from independent evaluations of computer use, coding quality, safety and total task cost. Customer disclosures will also matter: named deployments, production scale, failure rates and measurable time savings would provide stronger evidence than the early examples in OpenAI’s announcement.

Enterprise buyers should watch whether access remains limited to approved applications, how often confirmation policies interrupt workflows and whether administrators receive sufficient visibility into tool calls and data transfers. Researchers should look for methodology and reproducibility details behind the internal safety benchmark and the comparisons with GPT-5.6 Sol and Claude Fable 5.1.

The market will also reveal whether computer use becomes a practical complement to APIs or a fragile workaround for missing integrations. Adoption of the OpenAI API, ChatGPT Work and Codex in production will be a clearer test of Astra’s value than launch-day capability descriptions.

Creati.ai perspective

GPT-6 Astra is significant because OpenAI is presenting computer use as a deployment shortcut for business AI, not merely as a model capability. If the system can reliably work within existing applications while respecting permissions, it could reduce integration work for smaller teams and accelerate automation projects.

The evidence is not yet strong enough to confirm that outcome. The detailed claims come from OpenAI itself, and the available source material lacks independent benchmarks, named customer results and production reliability data. For now, Astra looks like an important test of whether AI agents can move from demonstrations into governed workplace execution.