On 3 September 2026, OpenAI launched GPT-6 Astra, presented as its “smartest and best-aligned” model (official announcement). Beyond the benchmark scores, the demonstration that matters to a creative studio comes down to three examples: a slide deck produced on OpenAI's own presentation template, a house modelled in Blender then made walkable in Unreal Engine 5, and websites hosted directly from ChatGPT. The model reaches Plus, Pro, Business and Enterprise subscribers over the coming days, and the API at $10 per million input tokens and $50 per million output tokens (The New Stack).
The facts
- The rollout starts on 3 September with a limited number of organisations, then extends “over the coming days” to ChatGPT Plus, Pro, Business and Enterprise subscribers, as well as to the API, Microsoft Azure and AWS Bedrock. Usage is included in existing subscriptions, with credits to buy beyond that; an Astra Pro version is reserved for the Pro, Business and Enterprise plans, and Enterprise administrators must enable the model, which is switched off by default at launch (OpenAI, 9to5Mac).
- In the API, the standard rate is $10 per million input tokens and $50 per million output tokens, with a Fast mode that is twice as quick and billed at twice the price. That is 2.5 times the promotional rate of GPT-5.6 Sol, at $4 and $20, according to Forbes, and the same pricing grid as Anthropic's Claude Fable 5.1 according to The New Stack.
- On the production side, OpenAI shows a slide deck about a fictional model, generated from a few slides of its own template, with the expected tone and layout; ChatGPT's Sites feature creates and hosts websites, web apps and games from a prompt. On BenchCAD, which scores the reconstruction of 3D objects as CAD code, Astra reaches 95.9% geometric overlap, against 83.3% for GPT-5.6 Sol and 84.3% reported for Claude Fable 5.1, OpenAI specifying that the Claude scores reflect three changes to the evaluation (OpenAI).
- On OSWorld 2.0, which measures computer use, the model scores 72.6% in around 40 minutes per task, against 65.7% in 75 minutes for GPT-5.6 Sol. Alex Mashrabov, co-founder of Higgsfield AI, says his most complex creative workflows run with up to 20% fewer tokens than the other models tested (OpenAI).
What does a model that respects a template change for creative production?
The most telling demonstration for a studio is the slide deck: a few slides from the in-house template as input, a complete deck as output, in the right tone and the right layout. OpenAI adds that the model was trained to retain only the context useful to the deliverable, leaving out superfluous information, and claims “stronger visual judgement” on the sites and renders it builds. Read from an art director's chair, this shifts the value to the input: the brand guidelines and the template become the machine's raw material, and a brand that has never formalised them will get a generic result however powerful the model.
The second example concerns space. Astra models a house in Blender then turns it into a walkable scene in Unreal Engine 5, “so that designers and clients can explore the volume before it is built”. A complete kart game, credited to designer Pietro Schirano, serves as the showcase for playable game generation. For a Luxembourg company, the transposition is concrete: a mock-up of a stand or a landing page in a few hours, to be validated before committing to production. Two precautions apply: the clips shown are edited excerpts, according to OpenAI's footnote, and the evaluation scores correspond to the model's maximum effort level, which increases latency and token consumption.
The launch itself was messy. Reuters, CNBC and The Verge published from 2.03pm New York time, working from the press kit, while the official page was only accessible at around 3.31pm, after going online and being pulled, according to the timeline reconstructed by Forbes. Greg Brockman, president of OpenAI, closed the press briefing with “Welcome to the AGI era”, while specifying that the term was now a “mission concept” rather than a contractual trigger. The New Stack notes that on DeepSWE, an agentic coding test, Meta's Muse Spark 1.3 claims 75.4% against 74.1% for Astra, and that the public leaderboard places Gemini 3.8 Flash and Claude Opus 5 at 74%: the lead on code remains disputed. OpenAI also acknowledges that Astra's written reasoning is harder to monitor than that of GPT-5.6 Sol, a point the company says it takes seriously. Aidan Clark, at OpenAI, puts the training run at more than 100,000 GPUs at the Stargate site in Texas, the largest the group has ever carried out.
AIxH's view
What an art director takes from this announcement is the place the template now holds in the chain. The model produces quickly from what it is given, and it produces accurately when the brand guidelines and reference formats already exist. That is the method we apply within the Content & social media service of AIxH, a social media agency in Luxembourg: human editorial direction, AI-accelerated production. Astra will make mock-ups and format variations cheaper to produce; it will also make the absence of a brand system more visible, because the output will look like everyone else's. The priority this autumn is therefore to document that system, presentation templates included, and to train the teams who will talk to the model, which our ChatGPT training in Luxembourg covers. The visual judgement OpenAI claims will be judged on the evidence, campaign by campaign.