GPT-6 Astra for Work: What the Launch Actually Changes

OpenAI introduced GPT-6 Astra as its most capable model for professional work, available in ChatGPT Work, Codex, and the API. The important change is not the higher version number. OpenAI is positioning Astra as a model that can work through the same applications people already use, including ones without an API, and it is pairing the launch with new admin controls for how far that access reaches. That shifts the adoption question from “is it smarter?” to “can we let it act inside our systems safely, and does it produce a better verified result on our own work?”

What changed

OpenAI calls Astra “the world’s most intelligent and aligned model” and says it is state of the art on computer use, browsing, professional work, software engineering, cybersecurity, and science. Those are the vendor’s claims. The more concrete changes are these.

It works through existing applications. OpenAI says that in ChatGPT Work and Codex, Astra can write code and operate the same applications people use every day, even when those applications lack an API. The pitch is that businesses can use it inside current workflows without extensive integration work first.

It has a price. API pricing starts at $10 per million input tokens and $50 per million output tokens. OpenAI argues the model completes tasks in fewer tokens with fewer retries, so cost per task can be lower than the token price suggests. On Terminal-Bench 4.0, OpenAI reports Astra at 57.9%, compared with 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1, at roughly 9% and 63% lower estimated API cost per task, respectively.

It ships with more controls. New enterprise admin controls let organizations restrict Astra to approved websites and desktop applications, manage uploads and downloads, and control browsing history. ChatGPT Work and Codex also include confirmation policies that can require approval before consequential actions, plus automated review of potentially unsafe or unauthorized tool calls.

It connects to more enterprise software. Alongside Astra, OpenAI launched enterprise plugins in ChatGPT Desktop from Oracle Analytics, Power BI, Navan, and Avalara, built on its browser-use capabilities.

It is off by default for enterprises. OpenAI says enterprise access is off at launch. Administrators enable it under their applicable rate card and agreement. Zero Data Retention is available for eligible API customers on supported endpoints, subject to approval.

Why it matters

A model that can operate software is a different risk category from one that only writes text. OpenAI acknowledges this directly. It says Astra is the first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework, and that it strengthened protections against both misuse and the model taking unauthorized actions.

OpenAI also reports results from an internal computer-use safety benchmark that tests scenarios such as exposing confidential information, sharing a dashboard too broadly, or deleting data. It says Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1. That is a meaningful claim, but it is an internal evaluation. Fewer unintended outcomes is not zero, and your environment will contain scenarios the benchmark did not.

Customer quotes in the announcement point to the qualities that matter most for knowledge work. Box said Astra was more than 10% less likely to make confidently incorrect assertions in its evaluation. Hebbia said it followed briefs 17% more faithfully than the next-best model and sourced claims to the right document 19% more often. These are early, vendor-selected testimonials, so read them as hypotheses worth testing rather than settled results.

For an individual, the useful test is whether Astra reduces revision cycles on real tasks. For a company, the risk is that a strong demo hides weak governance. A model that can take more actions needs narrower permissions, better logging, and clear approval boundaries.

What to test

Do not compare Astra with an older model using one polished prompt. Build a repeatable evaluation set from actual work: a research brief, a messy spreadsheet, a policy comparison, a coding task, and a document that must follow a house style. Our method for comparing AI models with a repeatable test lays out the scoring.

1. Choose ten representative tasks, including two known failure cases.
2. Remove confidential data unless the approved product and contract allow it.
3. Run the same inputs with your current model and Astra.
4. Require citations or inspectable evidence where the task depends on facts.
5. Record human review time and total cost per accepted result, not just generation time or token price.
6. Test how the system behaves when a source is missing or contradictory, and whether it declines to assert what the documents do not support.

If you plan to use computer use, add a second track:

• Start with the narrowest admin configuration: approved sites and applications only, uploads and downloads restricted.
• Turn on confirmation policies for any action that sends, shares, deletes, or pays.
• Seed test scenarios that mirror OpenAI’s own safety categories, such as an overly broad share or a destructive change, and check whether the confirmation step catches them.
• Confirm what is logged, who can see the logs, and whether you can reconstruct why an action happened.

What to verify before buying

Check current plan availability, regional access, data retention terms, Zero Data Retention eligibility, connector and plugin permissions, API pricing, and rate limits on OpenAI’s official pages. Confirm who in your organization can switch Astra on, since enterprise access starts disabled. Compare Astra’s per-task cost against cheaper tiers too. Our breakdown of GPT-5.6 Sol, Terra, and Luna shows how wide that price range already is.

What to watch next

Three things deserve attention. First, whether independent evaluations reproduce OpenAI’s Terminal-Bench and cost-per-task claims. Second, how the confirmation policies and automated tool-call review hold up in real deployments, especially given the Critical cybersecurity rating. Third, whether more enterprise plugins arrive, since the value of computer use depends on which systems it can reach under your controls.

The conclusion

GPT-6 Astra is OpenAI’s bid to move AI from drafting work to doing it inside business software. The launch pairs stronger claimed capability with real governance tools, and that pairing is the part teams should take seriously. Treat the announcement as the start of an evaluation, not the result. The best model is the one that improves the complete workflow under your constraints, with permissions you can explain and actions you can audit.