GPT-6 Astra

MA

Matt Wolfe9 Nov 2026

OpenAI model priced at $10/$50 per million tokens, slightly above Fable 5.2, and behind it on most benchmarks.

MA

Matthew Berman4 Nov 2026

Live demo

Described as OpenAI's best model, used in the paper to interpret and edit weights. It identified games and languages from raw weights, fixed a car-ramp network by editing 3.5% of the weights, wrote game-playing networks from scratch, and fixed looping in a small model to raise math scores.

MA

Matthew Berman29 Oct 2026

OpenAI model scoring 57 on the index, used as a comparison. It was previously top on Vending Bench at about $15,000.

MA

Matthew Berman21 Oct 2026

Berman says his editor started using Astra to edit videos, no longer touching the timeline, and they now produce more videos.

AI

AI Search6 Oct 2026

Creator says OpenAI's release page shows over 99% success on SREBench after four attempts, a benchmark of reverse engineering software from binaries. It is presented as the catalyst for AI decompiling games and software.

MA

Matthew Berman30 Sept 2026

OpenAI's frontier large model, priced at $10/M input and $50/M output, used as the comparison point for Sol. It performs slightly better at the top end of thinking tiers on OSWorld.

AI

AI Search24 Sept 2026

Used as comparison throughout; Opus 5.5 was judged better on piano, math video, Blender and game, while Astra's ray tracing looked slightly better. Astra is cheaper and hallucinates less, and beats Opus on Maze Bench.

MA

Matthew Berman23 Sept 2026

OpenAI's top frontier model, used as the comparison point in benchmarks. It remains the best performer but costs far more than Sol and Luna.

MA

Matthew Berman23 Sept 2026

Competing frontier model that Opus 5.5 is benchmarked against. It comes second on Terminal Bench and beats Opus 5.5 on Terminal Bench Science. The creator also calls it the best at creative writing.

AI

AI Search20 Sept 2026

GPT-6 Astra helped decrypt an 82-character German WWII Enigma message that had been unsolved for about 80 years. It searched archives, built its own Enigma simulator and ran experiments in parallel for roughly 10 hours, with a human choosing the target and guiding the work.

MA

Matthew Berman10 Sept 2026

Model OpenAI released the previous week that the creator says blew the world away. OpenAI said an internal model, still training, is significantly more capable than it and was used to solve the Navier-Stokes problem.

MA

Matt Wolfe9 Sept 2026

Live demo

New OpenAI model used to diagnose and rework the detector. It found flaws in the confidence logic and ran for 8+ hours without finishing. The creator says it and other top models can't reliably tell if a video is AI.

MA

Matthew Berman9 Sept 2026

Model released four days before the video; the creator says he and others with access are impressed. It is cited as the benchmark that OpenAI's next-gen model significantly surpasses.

AI

AI Search7 Sept 2026

Live demo

Creator tests GPT-6 Astra on ray-tracing physics sim, Unreal game, computer use (Photopea, sprite animation, piano), DAW music, Blender MCP, commercials, explainer video, vision and deep research. It scores top on many self-reported benchmarks and leaderboards (ARC AGI-3 over 60%, Frontier Math, OSWorld), but fails the frog test and gets 1/6 on brain tumors.

AI

AI Search6 Sept 2026

OpenAI's new model that uses the computer directly and posts state-of-the-art results on computer use, terminal coding, ARC-AGI-3 and FrontierMath. Rolled out to paid plans; the creator notes it ranks 2nd and 3rd on some independent indexes.

MA

Matthew Berman4 Sept 2026

Live demo

Creator got early access to GPT-6 Astra and demoed many one-prompt projects: a Fall Guys clone, an ASCII 3D city, a SimCity game built over a 5-day /goal run, 3D worlds, simulations and horror games. He also showed browser control, slide generation and noted weaknesses: AI-smell writing and repetitive forest-green flat design, though it is highly steerable.

MA

Matt Wolfe4 Sept 2026

Live demo

OpenAI model rolling out to limited organizations then Plus, Pro, Business and Enterprise users. It scored 74.1% on the coding benchmark and near-saturated ARC-AGI-3, and was first on his SVG bench. Its Megabonk clone was the best-looking.

MA

Matt Wolfe3 Sept 2026

Live demo

OpenAI's new flagship model, rolling out to limited organizations and then ChatGPT Plus, Pro, Business and Enterprise. The creator reviews benchmarks (ARC-AGI 3 99.9%, Terminal Bench Science 64.6%, DeepSWE 74.1%, 5th on Artificial Analysis) and tests it on SVG generation, a Mega Bonk clone, an interactive website, and computer-use tasks in Blender and Unreal Engine. He also shows community demos and finds it a bigger leap than recent models.

MA

Matthew Berman3 Sept 2026

Live demo

OpenAI's new frontier model, which the creator calls the best he has used. It leads most benchmarks shown (Arc AGI 3, FrontierMath Tier 4 at 97.6%, BenchCAD 95.9%, Exploit Bench 100%) and is strong at computer and browser use. Priced at $10/$50 per million tokens, with a fast mode at 2.5x speed for 2x price. Available on the OpenAI API, AWS Bedrock and Azure. Caveats: it tends to stop after about 30 minutes, has repetitive green and pastel design tendencies, and its writing still has an AI smell.

MA

Matt Wolfe3 Sept 2026

Upcoming OpenAI model announced as delayed for safety; it can find and exploit security vulnerabilities, reaching roughly 40% success versus GPT-5.6 Soul's 11.5% with fewer tokens. Release expected soon.

AI

AI Search9 Aug 2026

An internal version of OpenAI's upcoming model, codenamed Astra, solved or advanced 10 open math problems at about $2,000 of API-rate compute. Reasoning walkthroughs were published.

MA

Matt Wolfe7 Aug 2026

OpenAI post listing 10 math and theoretical computer science problems solved by Astra, likely a next-generation model not yet released.