Level 1 — Absolute Beginner

OpenAI is a computer company. It makes programs that can talk and write. These programs are called AI models.

On September 3, the company made a new model. Its name is GPT-6 Astra. The company says it is the best one they have ever made.

The company says the new model can use a computer by itself. It can also do many steps of work without stopping.

Some people at the company say this is a very big moment. But other people say it is not that big yet. The model is also not cheap to use.

computer

A machine that stores and works with information.

model

Here, a computer program that has learned from a lot of text.

company

A business that makes or sells something.

new

Made or started a short time ago.

best

Better than all the others.

step

One part of a job you do in order.

moment

A short period of time.

cheap

Costing little money.

Level 2 — Elementary

OpenAI released a new flagship model called GPT-6 Astra on Thursday, September 3. It replaces GPT-5.6 Sol as the company's most powerful system, and OpenAI describes it as the most intelligent and best-aligned model it has built so far.

The company says Astra is a generational leap in areas such as cybersecurity, professional work, software engineering and science. It is better at staying focused, keeping to the limits of a task, understanding what a user actually wants, and finishing long jobs that involve many steps.

Greg Brockman, the company's president, went further and suggested the model might eventually be seen as the arrival of artificial general intelligence, or AGI. OpenAI has previously defined AGI as a system that can do most economically valuable work as well as or better than a human, and it has not shown that Astra reaches that bar.

The rollout is happening in phases. Companies taking part in OpenAI's application-based cybersecurity programme get access first. On the standard interface the model costs $10 per million input tokens and $50 per million output tokens, about two and a half times the promotional rate for the older Sol model.

flagship

The most important or most advanced product a company offers.

aligned

In AI, behaving in the way its designers intend.

cybersecurity

The work of protecting computers and networks from attack.

generational leap

An improvement so large it feels like a whole new generation of technology.

rollout

The gradual process of making a product available.

phase

One stage in a process that happens in steps.

token

A small piece of text that AI systems count when measuring usage.

promotional rate

A temporary lower price offered to attract customers.

Level 3 — Intermediate

OpenAI released GPT-6 Astra on September 3, positioning it as a generational advance over GPT-5.6 Sol across cybersecurity, professional workflows, software engineering and scientific reasoning. The company describes the model as its most intelligent and best-aligned system to date and emphasises improvements in task adherence: staying inside the boundaries of an instruction, inferring what a user actually meant, and carrying multi-step work through to completion without drifting.

The most discussed technical detail is the model's use of what OpenAI calls opaque recurrence, in which a query is passed through the network repeatedly rather than being decomposed into written intermediate steps. That design is a departure from the visible chain-of-thought approach that dominated the previous generation, and it raises an obvious tension: performance gains obtained inside a loop that produces no legible trace are harder for outside researchers to audit.

The AGI framing has drawn the sharpest reaction. President Greg Brockman suggested Astra may eventually be regarded as the start of artificial general intelligence, but OpenAI's own published definition sets the bar at a system able to perform most economically valuable work at or above human level, and the company has not demonstrated that Astra clears it. The gap between the marketing language and the internal definition is not a small one, and several outlets highlighted it immediately.

Commercially the launch is deliberately narrow at first. Firms enrolled in OpenAI's application-based cybersecurity programme receive access ahead of the general population, and pricing on the standard interface is set at $10 per million input tokens and $50 per million output tokens, roughly two and a half times the promotional rate that had applied to Sol. That is a meaningful repricing, and it suggests the company expects the capability gain to justify a materially higher cost per task.

adherence

The act of sticking closely to a rule, plan or instruction.

infer

To work out something from evidence rather than being told it directly.

recurrence

The act of passing through the same process repeatedly.

decompose

To break something down into smaller parts.

legible

Able to be read and understood.

audit

To examine something carefully to check that it is correct.

framing

The way a subject is presented in order to shape how it is understood.

repricing

A significant change to what something costs.

Level 4 — Advanced

OpenAI's release of GPT-6 Astra on September 3 arrived wrapped in a rhetorical claim that the company's own documentation does not support. Greg Brockman's suggestion that the model may come to be seen as the onset of artificial general intelligence sits awkwardly beside OpenAI's published operationalisation of the term, which requires a system capable of performing most economically valuable work at or above human level. No evaluation released alongside the launch establishes that threshold, and the distance between the promotional register and the definitional one is wide enough that several outlets flagged it within hours.

The architectural news is more interesting than the branding. Astra employs what the company terms opaque recurrence: rather than externalising intermediate reasoning as text, the model routes a query through the network multiple times, accumulating computation in latent space. The efficiency argument is straightforward, since serialising reasoning into tokens is expensive and lossy. The interpretability cost is equally straightforward and considerably more consequential. Chain-of-thought traces, whatever their fidelity as an account of what a model is actually doing, at least furnished a surface that external researchers and red teams could inspect. Recurrent latent computation furnishes none.

That tension is sharpened by the sequencing of the rollout. Access is being granted first to organisations enrolled in OpenAI's application-based cybersecurity programme, which is to say that the earliest deployment of a model with reduced reasoning transparency is into the domain where offensive misuse is most immediately consequential. The company presumably regards its vetting process as sufficient compensation. Whether that judgement is correct is not something outside parties can currently evaluate, which is itself the point critics are making.

The economics are unambiguous. At $10 per million input tokens and $50 per million output, Astra prices at roughly two and a half times the promotional rate that applied to GPT-5.6 Sol, a repricing that only makes commercial sense if the marginal capability gain translates into materially higher willingness to pay per completed task rather than per token consumed. It is a bet that the market has moved from valuing cheap generation to valuing reliable completion, and the phased rollout is a reasonable way to test that proposition before exposing the full customer base to the new curve.

operationalisation

The act of turning an abstract concept into a specific, measurable definition.

register

The style or level of language used for a particular purpose or audience.

externalise

To express something internal in an outward, observable form.

latent space

The internal representation a model uses, not directly readable as text.

lossy

Discarding some information in the process of conversion or compression.