Video Summary - Did OpenAI actually build AGI? GPT-6 Astra first look

Tech/news summary (GPT-6 Astra and competing model releases)

Presented as “advanced” coding + knowledge-work models.
Two-model split by availability:
- Mythos: described as not usable
- Fable: described as the available one
Case study—debugging crash:
- A hedge fund used Fable 5.1 to analyze a program crash by using a memory snapshot.
- They located a crash address inside a compiled vendor library.
- Then they reconstructed/disassembled it to trace the issue to a vendor bug (no source access).
Biotech—protein target binding:
- Claimed improvement from ~10% to ~50% success in designing proteins that bind specific body targets.
Space/geoscience—Venus mapping:
- Used 30-year-old NASA radar data to train a neural network to build an elevation map for a region of Venus.

Framed as a frontier model and the fourth release in ~5 months.
Pricing:
- Standard endpoint: $1.25 in / $4.25 out
Contributor tier:
- 10¢ in / 20¢ out, but requires allowing Meta to train on user submissions (positioned as attractive due to cost tradeoffs).

Video focuses on whether OpenAI “actually built AGI,” using an “AGI if you got access” framing for selected influencers/enterprise.
Release-day disruption:
- An auto-generated claim said ChatGPT, Claude, Grok, and Cursor went down around the time of Astra coverage.
- Speculation ranged from a likely Azure outage to a more conspiratorial “kill the competition” theory.
Launch/embargo chaos:
- OpenAI launched a page; outlets published embargoed stories.
- OpenAI allegedly took down the announcement.
- Influencers competed on how long they’d had secret access.
- About 90 minutes later, Astra was said to not yet be public; rollout to Plus/Pro was expected over the next days.
- Sam Altman posted an apology; a “go to bed” response is quoted when asked whether to stay up for availability.
- Claim: the model underwent a formal review with the Trump administration prior to release.

Pre-trained on 100,000+ GPUs at Stargate (Texas).
Claimed distinction vs older OpenAI models:
- Earlier work involved more supervising during training
- Astra presumably shifts toward a different approach.

Can fill forms, crunch spreadsheets, and operate engineering tools including KiCad and Blender.

OSWorld benchmark:
- 73% score
- ~40 minutes per task
- Compared to Soul: 65% but ~75 minutes
Trust Me Bro benchmark:
- Claims: 100% on exploit bench, 65% on terminal bench
- Other notable results are contrasted with earlier Anthropic “bragging” metrics.
Arc AGI 3:
- 99%
- Presented as evidence of generalization rather than memorization
- Tied to discussion around “AGI” definition debates.

OpenAI’s claim: Astra reached a “critical cyber threshold” in a preparedness framework.
Framed capability:
- Able to find and exploit zero-days without a human specifying actions
- Video treats this as a major differentiator.

Astra pricing matched to Fable 5.1:
- $10 per 1M input tokens / $50 per 1M output tokens

Described as positive early-access reviews, typical of access groups (not loudly critical publicly).

Sherif Shamim:
- Recreated the Palace of Fine Arts (San Francisco) in Blender, described as nearly perfect.
OpenAI staff (Thomas Recart):
- Walked through a demo house—modeled in Blender and converted into a fully walkable Unreal Engine 5 scene.
Matt Schumer:
- Created a world in Unreal Engine populated with dozen-ish Astra agents.
- Reported immersive behavior (agents in his home “talking” about plans—comedic interpretation in the video).

Artificial Analysis – Independent Intelligence Index:
- Astra scored 61
- Described as exactly equal to GPT-5.6 Soul
- And 5 points behind Anthropic’s Fable 5.1
The video suggests the results don’t align with Astra’s strong benchmark/cyber claims.

Code Rabbit Security (sponsor) is mentioned as providing “continuous code security.”
Positioning:
- Uses reasoning-based agents rather than regex/brittle rules
- Agents are intended to think like an attacker and scan for vulnerabilities across a codebase
Prioritization factors:
- reachability, exploitability, blast radius
Outputs:
- plain-English explanations
- a suggested fix (approve/merge from a diff)
Claims:
- Reviews each PR before merge
- Supports scheduled deep scans
Offer:
- 10 free code scans