OpenAI Releases GPT-6 Astra Capable of Operating Blender
OpenAI has launched GPT-6 Astra, with paid users beginning to access it gradually. One of the official demonstrations involves building a 3D house in Blender, then importing it into Unreal Engine 5 to create a walkable scene. Early users have also released Three.js web 3D examples, such as a ship in a bottle, waves, and seagulls.
The model was previewed to a limited number of organizations on September 3, and the following day it was made available to ChatGPT Plus, Pro, Business, Enterprise, as well as API, Azure, and Bedrock users. The enterprise workspace is disabled by default and needs to be enabled by an administrator. The developer interface is priced at $10 per million input tokens and $50 per million output tokens, with additional credits available for purchase after exceeding the subscription limits. The public version will refuse certain advanced cybersecurity tasks, with full offensive and defensive capabilities remaining within the Daybreak trust circle.
The company highlights features such as computer operations: screenshots, page structures, and local software controls, allowing for the backend operation of Blender, browsers, and office suites. The geometric overlap rate for generating CAD code on BenchCAD is reported at 95.9%, compared to 83.3% for GPT-5.6 Sol and 84.3% for Claude Fable 5.1. Agents’ Last Exam reported 59.3%, higher than Sol's 53.6%. ARC-AGI-3 reported 99.9% under enhanced tool configurations, while FrontierMath Tier 4 reported about 98%.
The training scale is said to be the largest in the company's history, exceeding 100,000 GPUs, located in Texas Stargate. After an agent breach at Hugging Face in July, the company delayed the release and added security layers. President Greg Brockman told reporters that looking back at the timeline for general artificial intelligence, it may fall around this generation of models.
The 3D workflow remains a demonstration with users sharing images, not a one-click finished product factory. In comparative tests, both Astra and Fable 5.1 can create seaside houses, with Astra adding wrinkles, tableware, and distant plants; common web 3D features include physics, music, and completed videos, but screenshots cannot prove the maintainability of the project. Materials, topology, and modifiable project files still require human acceptance. Some developers have kept Astra solely for planning acceptance, delegating specific code modifications to the cheaper Luna.
In market mechanisms, what is purchased is computer-operated and 3D agency premium, while what is sold is a pure dialogue model. The computing bill shifts from chat to prolonged screen occupation. Beneficiaries are creators who have purchased Plus/Pro and installed the Blender and Three.js toolchain, as well as OpenAI, which receives high output unit prices for API; those under pressure are junior outsourcing that sells modeling by the hour, and shell products that only provide dialogue entry. Traffic is flowing towards models that can invoke local creative software, rather than just adding another drawing button.
Blender combined with Three.js lowers the barrier from idea to a rotatable scene, without eliminating material and project acceptance.
Source: Public Information
ABAB AI Insight
OpenAI has shifted the main battlefield of its models from answering questions to operating systems. In the GPT-4 era, the focus was on text, in GPT-5.6 it was on tool invocation, and now Astra is focused on continuous screen occupation. Blender and UE5 appear on the official website because the menus of 3D software are deep and the steps are long, which precisely tests whether agents can survive in a real GUI for an hour. Three.js sends the results to the browser, eliminating the need to install a game engine.
Capital is flowing from pre-training clusters to inference duration. The output token is $50 per million, and prolonged modeling will turn single tasks into small cloud bills. Creators use expensive models for planning and cheaper models for file modifications, using price tiers to combat high costs. Microsoft Azure and AWS Bedrock are synchronously listed, selling the same set of agent capabilities into enterprise procurement lists, rather than creating another consumer chat skin.
This aligns with Anthropic's computer usage and Claude Fable's 3D generation, as well as video tools that integrate generative models into Blender plugins. The industry is in a period of expansion from 'generating an image' to 'generating a modifiable project.' Game studios use gray boxes to have models produce playable prototypes, indicating that acceptance criteria have shifted from aesthetics to functionality.
Structural changes belong to technological substitution. The mechanism is that the professional barriers of 3D software lie in shortcuts and engine pipelines, not in aesthetic slogans; whoever can operate these software reliably will replace junior execution by the hour. Quality, topology, and delivery specifications still remain with humans, as clients modify files, not posters. As agents become stronger, billing units will shift from word count to minutes of computer usage.
ABAB News · Cognitive Laws
Models begin to operate software, and software becomes a true interface
Rotatable scenes are more valuable than flashy screenshots
Use expensive models for planning, and cheap models for file modifications