OpenAI Astra Will Drive the Browser. It Should Not Stamp the File.

Via OpenAI: GPT-6 Astra: A new generation of intelligence

OpenAI Astra is a computer-use model, not a transformation plan

OpenAI is rolling out GPT-6 Astra: a model trained to sit in the browser, fill forms, update a CRM, organize a calendar, and produce slides that match a template. The company calls it the world’s most intelligent and aligned model. For an owner-led firm, that sentence is a buying decision dressed as a product launch.

Astra is not a new system of record. It is a very capable intern for the hop — the stretch between the email that already arrived and the field that still has to be typed. OpenAI says it is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work. It reports saturating FrontierMath Tier 4 with a 98% score and ARC-AGI-3 with 99.9%. Those numbers are for the lab. The operating question is simpler: who is allowed to let it click, and who still has to stamp.

Plus, Pro, Business, and Enterprise users get it over the coming days, plus the API, Microsoft Azure, and AWS Bedrock. Usage sits inside existing subscription allowances, with extra credits for sale. Pro, Business, and Enterprise also get GPT-6 Astra Pro. Enterprise access is off by default. That last line is the part that should survive the demo.

Off by default is the tell

OpenAI did not hide the computer-use pitch. Astra can fill online forms, update customer records, draft in your editor, run frontend checks, install and test software, and troubleshoot what is on screen. On OSWorld 2.0, OpenAI reports 72.6% at roughly 40 minutes per task, versus 65.7% at roughly 75 minutes for GPT-5.6 Sol — about 47% less time. On Agents’ Last Exam, 59.3% versus 55.5% for Claude Opus 5.

Faster clicking is not a map of the firm. If the date still lives in the managing partner’s head, Astra will fill the wrong calendar faster. If the last version of the file has no convention, it will draft from the wrong one. Herbert Simon’s 1971 line still applies: a wealth of information creates a poverty of attention. A system earns its seat only if it absorbs more than it produces.

Alignment is the half professional firms should read twice. OpenAI built an evaluation informed by the Hugging Face incident: whether a model facing a hard or impossible task goes beyond its intended scope. It reports GPT-5.6 Sol did that 48% of the time without production safeguards; Astra did it in 0% of cases. Harvey’s Niko Grupen, quoted in the post, said Astra “distinguishes documents from established records.” That is the stamp. Keep it.

Then the cyber section. OpenAI says Astra meets the Critical threshold for cybersecurity under its Preparedness Framework. On ExploitBench, 100% versus 78.5% for Sol. During testing it found two previously unknown zero-days and says it is disclosing them to maintainers. The launch build will refuse advanced tasks such as creating proof-of-concept exploits; a later Daybreak track is planned with looser safeguards for defensive work. Extra checks can pause or stop a task. In ChatGPT or Codex you may have to review; in the API the task stops.

A model that can drive a browser and also sit at that capability level is not a toy you turn on for the whole office because the slides look good.

API list price: $10 per million input tokens, $50 per million output. Fast mode: up to 2× speed at 2× the standard price. Zero Data Retention for eligible API customers. Staff on personal Plus seats will not wait for the Enterprise toggle.

What smart firms do before they enable Astra

Map the stream before you let a computer-use model click.

1. Inventory who already has Plus or Pro. Enterprise off-by-default does not stop the personal account.
2. Name the machines Astra would touch. CRM. Calendar. Inbox. Document editor. What is lost if a click is wrong.
3. Decide the stamp. The model may prepare a form, a recap, a books feed. A person accepts. Mail to clients stays human.
4. Put a convention on folders before you ground anything. Computer use on an unmapped drive is still an unmapped drive, only faster.
5. Price the meter. $10 / $50 and fast-mode doubling are not “included” once credits start. Someone has to watch spend.
6. Keep computer use inside the suite you already run. Do not add a third dashboard because the demo filled a form.

Rejected on the first pass: turning Astra on for the whole workspace because the benchmark chart is pretty. AI-drafted client correspondence. Asking the model for exploit work the launch build is already set to refuse.

How we actually run this work

This is a Workflow ROI Audit, not a model bake-off.

We map where expensive time actually goes — search, re-key, follow-up, the hop into the CRM — and where a computer-use model would save money versus where it would only speed up the wrong click. You get a systems map you can act on with us or without us.

If confidentiality is the constraint, pair it with a Shadow-AI Risk Assessment. Plus and Pro will be on phones before the admin finds the Enterprise toggle. Know what is leaving the firm.

Then Managed AI Operations: who is allowed to enable Astra, whether pauses and reviews are actually on, whether unused seats and extra credits are leaking, and whether the file got better. Models change. The memory of how the firm works should stay yours.

We do not rip out the suite you already paid for. We do not recommend Astra because it won a chart. We do not build against an open meter.

The paid entry is a Workflow ROI Audit when you want the costed roadmap first.

Let Astra prepare. Keep the stamp.

If your plan is “turn on GPT-6 Astra and let it drive,” you are automating the old path with a better intern. OpenAI left Enterprise off by default. Take the hint.

Ask a narrower question: which hop is worth a computer-use model, who may click, and who still stamps?

That is how OpenAI Astra pays. Map first. Let it prepare. Keep the stamp.

This article summarizes publicly reported information and is for general informational purposes only. It does not constitute legal, tax, financial, investment, security, or compliance advice. BuildBrain is not a law firm, accounting firm, or registered investment adviser. Facts, pricing, statistics, and product capabilities cited here reflect the sources listed at the time of writing and may change. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances—including applicable federal, state and local requirements. BuildBrain may have commercial relationships with vendors mentioned; where material, such relationships are disclosed. Nothing in this article is an endorsement of any specific AI product, model, or provider.