GPT-6 Astra: Computer Use, Pricing, and Alternatives (2026)
GPT-6 Astra is OpenAI's computer-use model, announced on September 3, 2026. It operates software and browsers directly, and OpenAI says it produces polished documents, spreadsheets, and presentations. API pricing is $10 per million input tokens and $50 per million output tokens.
What OpenAI announced
The announcement went live on September 3, 2026. OpenAI describes GPT-6 Astra as "the world's most intelligent and aligned model." The post calls it "state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work." Those are six distinct areas, and computer use is the one that changes daily workflows.
Rollout is staged. The launch post says Astra "is rolling out today to a limited set of organizations." Over the following days it becomes available to all ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API and AWS.
Three benchmark families anchor the claim: agentic computer use, scientific workflows, and reasoning. The gating has a stated reason. OpenAI writes that Astra "meets the Critical threshold in cybersecurity under our Preparedness Framework." The company notes that the same skill that finds zero-day exploits also helps defenders patch them.
Benchmark results published with the launch include:
| Benchmark | GPT-6 Astra | Comparison figure OpenAI published |
|---|---|---|
| ARC-AGI-3 | 99.9% | Described as saturated |
| FrontierMath Tier 4 (v2) | 98% | Described as saturated |
| ExploitBench | 100% | Described as saturated |
| Terminal-Bench Science 0.1 | 64.6% | 52.6% for Claude Fable 5.1 |
| OSWorld 2.0 (offline set, partial score) | 72.6% at roughly 40 minutes per task | 65.7% at roughly 75 minutes |
| ScreenSpot-Pro (no tools) | 92.7% | 76.9% for GPT-5.6 Sol |
Why this matters if your job ends in a file
Most model launches are interesting to read and hard to act on. A benchmark table rarely tells you whether Thursday gets easier. This one carries a line that does.
OpenAI writes that Astra "combines the intelligence required for complex problems" with the ability to run multistep workflows. The same sentence says the result is "polished documents, spreadsheets, and presentations." On slides the post is more specific. It calls Astra "our best model for adhering to existing templates." The slides it produces are described as "well laid out" and as conveying "key points with a structured narrative."
There is a second claim worth noting. The post says Astra "creates clear, well-structured documents, presentations, spreadsheets, and analyses that follow your templates and match your writing and visual style." It is also "trained to specifically pull only the context that matters into outputs, instead of repeating information unnecessary for the work at hand."
OpenAI illustrates this with a slideshow built from a few slides of its own presentation template, described as "capturing the correct tone and layout throughout." The computer-use examples span circuit board layout, an Excel competition, filling in a Form 1040, Power BI, and formatting a legal document.
Pricing as published
| Item | Published figure |
|---|---|
| API standard input | $10 per million tokens |
| API standard output | $50 per million tokens |
| Fast mode | Up to 2x the speed of Standard processing at 2x the Standard price |
| Cache reads and writes | Separate rates apply |
| Model ID | gpt-6-astra |
| Also available through | Amazon Bedrock |
The API figures are the only prices OpenAI published for the model itself. Consumer access rides on existing ChatGPT plans rather than a separate line item. The launch post lists Plus, Pro, Business, and Enterprise as the tiers that receive it.
Which task suits which route
It is worth being concrete about when each approach earns its keep. A computer-use model and a data agent solve different halves of the same day. The split below is about where your work starts, not about which product is better.
| If your starting point is | The useful route |
|---|---|
| An app with no API, and a task a person would click through | A computer-use model driving the interface |
| A messy export that nobody has cleaned yet | A data agent that reads the file directly |
| A template your company already uses | A model that adheres to templates, as Astra's post describes |
| A recurring report on the same schema each month | A saved workflow you can rerun without re-explaining it |
Where a stronger model stops being the bottleneck
There is a limit that no model release removes. If the underlying export is ambiguous, a better model produces a more confident version of the same ambiguity.
Three examples come up constantly. A revenue column that mixes gross and net. A date column stored as text in two formats. A customer ID that changed halfway through the year.
None of these are reasoning problems. They are decisions about your business, and somebody has to make each one and write it down. A model that guesses well still guesses. The fix is upstream of the model, in the file itself.
This is the layer where Powerdrill Bloom sits. It works from the file rather than from the screen. You upload the export and it reads the columns. One pass produces both the analysis and the deliverable. The distinction is not capability, it is starting point. Driving an interface and reading a file are different jobs, and plenty of teams need both.
Alternatives worth knowing
- Claude Cowork. Anthropic's product page says it "completes tasks you can steer from anywhere" and "works directly in folders and tools you choose." See what Claude Cowork is for the longer definition.
- Google Pics. For image work inside Workspace rather than general computer use. Covered in the Google Pics rundown.
- A file-first data agent. If your inputs are spreadsheets rather than screens, the AI report generator and Excel to PPT routes skip the interface layer entirely.
Conclusion
GPT-6 Astra is a genuine step in operating software, and the template-adherence claim is the part most likely to change a working day. Access is staged and cybersecurity-gated, so what you can use this week depends on your plan.
If your bottleneck is the export rather than the interface, start from the data instead. Try Powerdrill Bloom and see what one file produces.
Frequently asked questions
When was GPT-6 Astra released?
OpenAI published the GPT-6 Astra announcement on September 3, 2026. The safety overview carries the same date. The post describes a staged rollout that begins with a limited set of organizations.
How much does GPT-6 Astra cost through the API?
OpenAI lists $10 per million input tokens and $50 per million output tokens for standard processing. Fast mode runs at up to twice the speed for twice the standard price. Separate rates apply to cache reads and writes.
Can I use GPT-6 Astra on a ChatGPT plan?
The launch post says Astra becomes available to all ChatGPT Plus, Pro, Business, and Enterprise users over the days following launch. It is also offered through the OpenAI API and AWS.
Why is access to GPT-6 Astra restricted at launch?
OpenAI states that Astra meets the Critical threshold in cybersecurity under its Preparedness Framework. The company links the restriction to that assessment rather than to capacity.
Does GPT-6 Astra replace a data analysis tool?
They start in different places. Astra is described as operating software and browsers, while a data agent reads your exported file directly. Teams with both messy spreadsheets and click-heavy apps often use one for each.