GPT-6 Astra is out: what it means for business in Uzbekistan

GPT-6 Astra is OpenAI's new flagship model. Limited access opened on 3 September, general access on 4 September. The previous flagship was GPT-5.6 Sol, released in July.

The plain numbers: the model holds 1,050,000 tokens of context, writes up to 128,000 tokens per answer, and its knowledge runs to 30 April 2026. It was trained on more than 100,000 GPUs at the Stargate facility in Texas.

The practical difference comes down to one word: autonomy. Earlier models answered a question. Astra is built to take a task, work through several steps and deliver a result.

This is the part that matters to us. A lot of time in automation projects goes into the same situation: a system has no API, so an employee moves data from one window to another by hand every day. Until now that was hard to automate.

This is exactly where Astra jumped. The figures OpenAI published: on OSWorld 2.0, which measures real work done on a computer, it scores 72.6 percent against 65.7 for the previous model. The score matters less than the clock: Astra spends about 40 minutes per task where the previous model spent 75.

On finding an element on screen (ScreenSpot-Pro) the gap is wider still: 92.7 against 76.9 percent. On terminal work it scores 57.7 against 37.3. The model can now look at an interface, find the right control and act in sequence.

One more improvement: it makes things up less often. On OpenAI's internal measure, the share of factual errors fell from 12.2 to 4.2 percent.

This part cannot be skipped, because it affects the decision to use the model at all.

Astra uses a new technique called recurrent depth. In plain terms, part of the model's reasoning now happens in an internal state rather than in words. As a result you cannot read from the outside why it reached a given decision.

This was not alleged by critics; OpenAI stated it. Its official safety document says the ability to monitor the chain of thought has decreased substantially compared with earlier models, and that if the model deliberately underperformed and hid it, they would likely be unable to catch it.

Second, Astra is the first model in OpenAI's history rated at the top critical level for cybersecurity. During testing it found two previously unknown vulnerabilities. For that reason the capability is off by default for enterprise customers and an administrator must switch it on.

Care is needed here. OpenAI called it the most intelligent model in the world, but that claim comes from its own tables.

The independent Artificial Analysis index tells a different story: in the latest measurement Astra and Anthropic's Claude Fable 5.1 scored level. Astra leads on some tests (terminal work, automation) and trails on others (scientific code, long-context understanding).

On a few benchmarks Astra came out below even its predecessor. On Humanity's Last Exam it scored 57.2 percent against 65 for Claude Fable 5.1. OpenAI published that result rather than hiding it.

Independent testers noted something else: Astra writes worse prose than the previous model. The style is dry and impersonal, and some reviewers advised against using it for creative work at all.

There is also a speed problem: Astra is slow to answer. The first word can take up to a minute and a half. For a chatbot answering a customer in real time, that does not work.

First, cost. Astra is 2.5 times more expensive than the previous model. If your bot answers a simple question a thousand times a day, Astra multiplies the bill and adds no value at all. A cheap model is enough for a simple task.

Second, the place where Astra genuinely pays off is specific: working with legacy systems. Many enterprises in Uzbekistan run software with no API, and staff carry data between those programs by hand. Astra can start taking that work over.

Third, the security angle. If your data is sensitive, the fact that the model's reasoning cannot be inspected is a serious question. On such a project you settle how the model will be supervised before you choose it.

Fourth, and most important: one model is not the answer to everything. A project runs several models. Astra handles the complex steps, a cheap model handles simple replies, a third writes text. That is how we build client projects, because it is both cheaper and better.

Our rule is simple: the task decides, not the brand. At the start of a project we split the work into parts and put the cheapest sufficient model on each part.

In practice: a fast, inexpensive model in the customer-facing chatbot, a mid-tier model for reporting and analysis, and an Astra-class model for multi-step autonomous work. That keeps the bill down and the quality up.

A new model shipping does not mean everything should move to it. We run it through our own tests first, and only then put it into a client project.