AI Model Releases: A Chaotic Week

The video reviews three major AI releases and the unusually messy rollout of OpenAI’s GPT-6 Astra. It also questions whether Astra’s impressive demos match its independent benchmark results.

Anthropic: Fable and Mythos 5.1

• Fable is available to users; Mythos is restricted and may avoid certain sensitive topics.
• A case study showed Fable diagnosing a rare crash by examining a memory snapshot, disassembling a proprietary vendor library, and tracing the fault to a bug in that library.
• Mythos reportedly improved AI-designed protein candidates’ success rate from roughly 10% to 50%.
• It also used decades-old NASA radar data to produce a new elevation map of part of Venus.

Meta: Muse Spark 1.3

• The fourth release from Meta Superintelligence Labs in five months.
• Its standard pricing is $1.25 per million input tokens and $4.25 per million output tokens.
• A much cheaper contributor tier costs $0.10 in / $0.20 out, but allows Meta to train on submitted data. Meta says a double-digit share of developers have opted in.

OpenAI: GPT-6 Astra

Rollout

• ChatGPT, Claude, Grok, and Cursor experienced outages around the launch; an Azure issue was considered the likeliest explanation, though the timing prompted jokes about Astra eliminating rivals.
• OpenAI briefly removed its announcement after media stories appeared. Astra was not immediately available to the public; Plus and Pro access was expected to roll out later.
• Sam Altman apologized for the launch confusion and said the model had undergone a formal review with the Trump administration.

Capabilities and benchmarks

• OpenAI says Astra was pretrained using more than 100,000 GPUs at its Texas Stargate site. It is also the company’s first model in which earlier models performed a substantial share of training supervision.
• Its main focus is computer use: filling forms, working with spreadsheets, and using tools such as KiCad and Blender.
• Reported results include:
• 73% on OSWorld, with tasks averaging about 40 minutes; the video compares this with Soul’s 65% at 75 minutes.
• 100% on Exploit Bench and 65% on Terminal Bench Science.
• 99% on ARC-AGI-3.
• OpenAI says Astra crossed its preparedness framework’s critical cyber threshold, meaning it can autonomously find and exploit zero-day vulnerabilities.
• Pricing matches Fable 5.1: $10 per million input tokens and $50 per million output tokens.

Early demonstrations and mixed results

• Early-access feedback was generally positive, though the video cautions that such reviews may be biased.
• Demos showcased Astra recreating San Francisco’s Palace of Fine Arts in Blender, converting a Blender house model into a walkable Unreal Engine 5 scene, and populating an Unreal world with AI agents.
• Despite the demos, Astra scored 61 on Artificial Analysis’s Independent Intelligence Index—equal to GPT-5.6 Soul and five points below Fable 5.1. The host says this discrepancy raises questions about the claims and comparisons.

Sponsor: CodeRabbit Security

• The video promotes CodeRabbit Security, a tool that uses reasoning-based agents to review code for vulnerabilities.
• It evaluates risks such as reachability, exploitability, and blast radius, explains findings, and suggests fixes.
• It reviews pull requests and supports scheduled full-codebase scans. The sponsor offered 10 free code scans.