OpenAI announces GPT-6 Astra and one billion in advertising revenue: figures stated by the company

On September 4, 2026, OpenAI announced GPT-6 Astra, a model that the company says will be rolled out across all ChatGPT plans in the following days. According to the company, the model was pre-trained on more than 100,000 graphics processing units within the Stargate infrastructure. The announcement is covered in the roundup by Tech Startups and in the newsletter of Crypto Integrated, which places alongside the company’s statements the evaluations published by ARC Prize.

The two write-ups, however, share the same primary source: the company’s press release. So the boundary of what is confirmed must be set right away.

What is claimed and by whom

OpenAI attributes to Astra a score of 63% on the ARC-AGI-3 benchmark and performance above the human baseline on 96% of the levels. These are figures communicated by the model’s maker: in the material available, there is no measurement conducted and published by an independent third party on the announced version. The distinction is not a formality. A benchmark measures a system’s behavior on a defined set of tasks; the result depends on how the model is queried, how many times, and with what computing resources. If the party making the claim is the one selling the system, the number remains a statement from that party until someone else replicates the measurement.

The 63% score on ARC-AGI-3 is communicated by OpenAI: in the material available there is no published independent verification of the announced version.

On the same day the company presented “Defense Factory,” described as an automated system for identifying, validating and fixing cybersecurity vulnerabilities. Technology roundups from September 4 also note a motion in the United States Congress to survey out-of-control artificial intelligence agents: two elements that, taken together, indicate where the regulatory debate is shifting, from generated content to the autonomous capacity for action of these systems.

The second announcement: advertising

The economic angle is covered by the industry newsletter AI to ROI, which reports statements from OpenAI and data communicated by Criteo. On this matter the news currently comes from a single source; no independent confirmation is available in the editorial team’s material.

OpenAI states that advertising within ChatGPT has reached an annualized revenue rate of $1 billion, about 200 days after launch. How the figure is constructed matters: annualizing means projecting a recent period’s trend over twelve months, and it is not the same as a finalized annual revenue figure. The same source notes that the figure remains below the $2.4 billion the company had projected as advertising revenue for the whole of 2026.

On the operational side, ad testing had begun in February; the company is now rolling out direct purchase of ad space by advertisers in India, Europe, the Middle East and North Africa. Criteo, named as the program’s first technology partner, reports more than 2,000 brands already active through the integration: this too is a figure communicated by a company that takes part in the program.

What we don’t know

We do not know the methodology used to calculate the ARC-AGI-3 score, nor the model’s energy consumption or operating cost. We do not know how many users are already seeing the ads, in what placement within responses, or how revenue attributed to advertising is accounted for relative to subscriptions.

The first genuinely verifiable element will be the extension of direct ad-space purchasing to European markets: that will be visible from the outside, with no need for a press release.