GPT-6 Astra Is Finally Here (And It’s REALLY Good)

Summary
Matt Wolfe presents an overview and hands-on testing of OpenAI’s newly launched GPT-6 Astra model on its release day (September 3, 2026). He breaks down the official benchmark results, tests the model on custom coding, SVG generation, and 3D web app tasks, and showcases its native computer-use capabilities by having it autonomously model, rig, and animate a character in Blender before embedding it into a playable Unreal Engine environment.

---

What is shown

• Announcement and Blog Post Overview [01:00]: OpenAI’s release page detailing GPT-6 Astra rollout plans to ChatGPT Plus, Pro, Business, Enterprise, and API users, along with demos such as circuit board design, Excel financial modeling, and tax form automation.
• Official & Third-Party Benchmarks [01:48]: Tables and graphs showing results across AutomationBench, DeepSWE (74.1%), Terminal-Bench Science, ARC-AGI-3 (99.9%), and the Artificial Analysis Intelligence Index and cost rankings ($1.67 per task).
• BuseyBench Test [05:27]: Evaluation of GPT-6 Astra generating raw SVG code for a portrait of Gary Busey, taking 1st place with an AI judge score of 7.1, using 63,858 tokens over 9 minutes.
• 3D Web Game Generation (Megabonk Clone) [07:32]: GPT-6 Astra writes a Three.js-based 3D action game in 8 minutes, featuring selectable classes (Mage, Ranger, Knight), procedural terrain, enemies, and leveling upgrades.
• Interactive Web Showcase (“Orbis”) [09:29]: An interactive 3D planetary simulation website built via Codex prompt in under 2 minutes, featuring real-time sliders for sunlight, sea level, rainfall, clickable terraforming tools, asteroid impacts, and population/habitability telemetry.
• Autonomous Computer Use in Blender & Unreal Engine [11:33]:
  • Prompting GPT-6 Astra to control Blender to sculpt a 3D humanoid wolf figure [11:45].
  • Prompting the model to rig the model with a 50-bone skeleton and animate an idle animation [12:17], followed by a running loop [12:38].
  • Taking control of Unreal Engine to construct an interactive forest environment (“Whisperwood”) and import the animated wolf as a playable character controlled via WASD [13:23].
• Community Creations on X [15:00]: Demos from Matt Berman, Matt Shumer, Pietro Schirano, and Ethan Mollick showcasing 3D planet explorers, an Unreal Engine Manhattan rebuild, an autonomous multi-agent settlement, a Fall Guys clone, jet ski racing, and an interactive historical simulation of the Library of Alexandria.

---

Claims & numbers

• Availability: GPT-6 Astra began rolling out on September 3, 2026 to select organizations, with general access for ChatGPT paid tiers and API arriving over the following days (the presenter states).
• AutomationBench: Jumped from 18.1% to 41.0% (the presenter states).
• DeepSWE: Scored 74.1%, marginally ahead of GPT-5.6 Sol (72.1%), Gemini 3.8 Flash (74%), and Claude Opus 5 (74%), but trailing Meta’s Muse Spark 1.3 (75.4%) (the presenter states).
• Terminal-Bench Science 0.1: Improved from 22.0% to 64.6% (the presenter states).
• ARC-AGI-3: Reached 99.9% with provider harness (compared to 7.8% for prior models and a 48% human tester average) (the presenter states).
• Artificial Analysis Intelligence Index: GPT-6 Astra placed 5th on the composite index, essentially tied with GPT-5.6 Sol around 68–69 points (the presenter states).
• Cost: Ranked at approximately $1.67 per task on the Artificial Analysis index, slightly higher than GPT-5.6 Sol but cheaper than Claude Fable 5.1 ($6.87 per task) (the presenter states).
• Generation Times: The BuseyBench generation took 9 minutes; the 3D Megabonk game took 8 minutes; the initial Blender character sculpt took 8 minutes; the 50-bone rigging and idle animation took 6 minutes; and the Unreal Engine environment setup took 35 minutes (the presenter states).

---

Notable quotes

• [00:10]: “This launch feels too big to just be like a footnote in an AI news video.”
• [03:44]: “It went from 7.8% to 99.9%... It’s basically solved this benchmark.”
• [12:54]: “It did this by taking control of my computer, opening Blender, drawing out the wolf for me, and then after that it rigged it up with 50 different bones... all through prompts.”

---

Assessment
This is a hands-on independent review and demo video by creator Matt Wolfe evaluating early access to GPT-6 Astra. The demonstrated prompts, Codex interfaces, and resulting 3D outputs (in-browser Three.js and live screen-capture of Blender and Unreal Engine) are shown directly running on his system, with honest commentary noting minor graphical quirks and movement glitches alongside high overall capability.

Described by gemini-3.8-flash on 2026-10-19 from the video’s audio and frames.