GPT-6 Astra is OpenAI's new model for complex work involving reasoning, software, research and computer use. For a business owner, the useful question is whether it can help finish a specific job: improve a service page, investigate a broken enquiry form or prepare a website update that someone can review.
This guide explains the confirmed details and how we would approach a first trial for a Kenyan business website.
What is GPT-6 Astra?
OpenAI released GPT-6 Astra on 3 September 2026. The company describes improvements in browsing, computer use, coding and professional document creation. Its examples include building websites and checking frontend functionality. These are capabilities described by the vendor, rather than results from a DoWebsites test. OpenAI's announcement
Astra is a model. What it can do in your workflow also depends on the application hosting it, the tools connected to that application and the permissions you provide. Giving a chatbot a website address is different from giving a coding assistant access to the site's project files.
For a useful brief, specify the result, the available materials and how success will be checked.
What has improved compared with GPT-5.6 Sol?
OpenAI reports the following results:
Evaluation | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
Terminal-Bench 4.0 | 57.9% | 37.3% |
OSWorld 2.0 | 72.6% | 65.7% |
DeepSWE v1.1 | 74.1% | 72.7% |
These figures come from OpenAI's launch comparison. Its notes say evaluation environments can differ from production ChatGPT. The size of the gain also varies by benchmark. Benchmark results and methodology notes
Treat those numbers as a reason to test a relevant task. They do not tell you how many corrections your own website will need or what a completed job will cost.
GPT-6 Astra pricing and availability
At the time of research, OpenAI describes a staged rollout beginning with selected organizations, followed by access through ChatGPT Plus, Pro, Business and Enterprise, and the API. Check your account before planning work around availability. Rollout details
The API documentation lists these specifications:
Item | Published specification |
|---|---|
Model identifier |
|
Context window | 1,050,000 tokens |
Maximum output | 128,000 tokens |
Standard input price | US$10 per million tokens |
Standard output price | US$50 per million tokens |
These are standard API token rates, not ChatGPT subscription prices or a fixed charge per website. Caching and tool charges can change the bill. Official model documentation
For illustration, 100,000 standard input tokens and 10,000 billed output tokens would cost US$1.50 in those token charges: US$1.00 for input plus US$0.50 for output. This is arithmetic, not an estimate of a typical Astra task. A workflow may make multiple requests.
For a budget in Kenyan shillings, convert the actual dollar charge using your payment provider's rate and include applicable charges. We have not assumed a fixed exchange rate or a Kenya-specific subscription price.
Four ways to evaluate Astra on a business website
The following are suggested trials, not claims that DoWebsites has measured these outcomes.
1. Turn a vague service page into a useful brief
Suppose a Nairobi installation business has a page that says little beyond “quality solutions.” Give the assistant an approved list of services, service areas, customer questions and genuine project examples.
Ask it to propose a page structure that helps a visitor establish whether the business handles their job. Require it to mark missing facts instead of inventing prices, qualifications or testimonials.
Review the result against the information your sales team actually uses. A useful draft should reduce the questions a qualified customer has to ask before enquiring.
2. Investigate a broken enquiry journey
A complaint such as “the website does not bring leads” is too broad for a first technical task. Start with a reproducible problem: a form fails on mobile, a button points to the wrong number or the success message appears without an enquiry arriving.
Provide a test environment and a precise expected result. Ask for the cause, the proposed change and evidence from repeating the original steps.
Check both ends of the journey. A visible success message alone is not proof that your team received the enquiry.
3. Prepare a website refresh
Ask for an inventory of the current pages, then identify which ones need updated copy, navigation changes or technical repairs. Supply your current business priorities so the proposed work has a clear purpose.
The useful deliverable is a reviewable list of changes, with a reason for each. A new visual style is only one possible outcome.
If you have not decided how much work the site needs, start with our guide to when to repair, refresh or rebuild a website.
4. Explain a supplied performance report
Give the assistant a small, clearly labelled export and ask it to explain what the figures show. Include the reporting period, the meaning of each column and any known tracking changes.
Require it to separate observations from possible explanations. If enquiries fell while traffic stayed steady, that identifies a question to investigate; it does not establish the cause.
Check its calculations against the source data before using the report to change your budget.
A prompt for your first trial
Use a bounded task that you can inspect:
Review the mobile enquiry journey on our staging website. Test the service-page enquiry button, form validation and confirmation state using dummy data. Record the steps and evidence for each issue. Propose fixes and identify anything you could not verify. Do not publish changes or contact real customers. Success means a test visitor can complete an enquiry and we can verify its receipt.
Adapt the scope to the tools available. If the assistant cannot open a browser or access the receiving system, it should report that gap. You can then supply the missing evidence or complete that check yourself.
What still needs human review?
OpenAI's safety overview places Astra at the Critical cybersecurity capability level under its Preparedness Framework. It reports better resistance to prompt injection and fewer unauthorized actions in evaluations, while acknowledging challenges monitoring the model's reasoning under adversarial conditions. That is not a guarantee of error-free operation. OpenAI's safety overview
For website work, our recommendation is to review changes wherever a mistake affects a customer or business record. Confirm prices, contact details and service claims. Use dummy data for test submissions. Check the exact page and interaction that changed before releasing it.
A successful demonstration is a starting point for this review, not a substitute for it.
Is GPT-6 Astra worth trying?
Our editorial view is that a focused trial is more useful than rebuilding your workflow around the launch.
Choose one recurring job, keep the brief and record four things: time to completion, usage cost, corrections required and whether the final result passes your checks. Include your own review time. Repeat with comparable work before drawing a conclusion.
For a DoWebsites reader, the outcome worth paying for is a better working website: clearer service information, a completed customer journey or an update your team can maintain.
Frequently asked questions
Is GPT-6 Astra the same as a website builder?
The model can be part of a website-building workflow. Your actual editing, testing and publishing options depend on the application and tools around it. Confirm those capabilities before giving it a build brief.
Does a large context window mean I should upload everything?
Start with the material needed for the task. An approved service brief and a few relevant pages may be easier to check than a large collection of outdated files.
Will using Astra improve my Google rankings?
This article establishes no ranking benefit from using Astra. Evaluate the accuracy and usefulness of the resulting pages, and measure performance after changes. Do not treat the model name as evidence of SEO quality.