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

Summary
In this episode of The Code Report, host Jeff Delaney rounds up a rapid-fire week of major frontier AI releases in early September 2026, highlighted by Anthropic’s Claude Fable 5.1 and Mythos 5.1, Meta’s Muse Spark 1.3, and OpenAI’s GPT-6 Astra. He recaps the chaotic rollout of GPT-6 Astra, its reported benchmark leaps, early access demos in 3D modeling and agent simulation, and contrasting independent evaluation results.

What is shown

• [00:11 - 02:23] Anthropic Claude Fable 5.1 & Mythos 5.1: Promotional materials and case studies showing Claude Fable 5.1 resolving a 5-year-old one-in-a-million crash for hedge fund Millennium by disassembling vendor code into assembly, Mythos 5.1 designing protein binders (Nipah G, EGFR, GDF-8), and generating Venusian elevation maps from 30-year-old NASA radar imagery.
• [02:24 - 03:03] Meta Muse Spark 1.3: Presentation of benchmark charts, pricing tiers ($1.25/$4.25 standard vs. $0.10/$0.20 contributor tier), and quotes from Mark Zuckerberg and Alexandr Wang.
• [03:04 - 04:19] The Outage and Launch Rollout: Coverage of the simultaneous multi-platform outage across ChatGPT, Claude, Grok, and Cursor; the premature launch page, 404 retraction, media embargo breaks (CNBC, The Verge), influencer early-access posts on X, and Sam Altman telling users to "go to bed."
• [04:20 - 05:36] GPT-6 Astra Capabilities & Benchmarks: Overview of Astra’s training on 100,000+ GPUs at the Texas Stargate site; demonstrations on OSWorld (filling Form 1040, spreadsheet puzzles, FreeCAD 3D transmission modeling); benchmark comparisons across ExploitBench, Terminal-Bench Science, FrontierMath Tier 4, and ARC-AGI-3.
• [05:37 - 06:36] Early Access Demos & Evaluations: Demos including Sharif Shameem’s Blender reconstruction of San Francisco's Palace of Fine Arts, Thomas Ricouard’s walkthrough of an Astra-modeled house in Unreal Engine 5, Matt Shumer’s simulation of talking Astra agents in Unreal Engine, and the Artificial Analysis Intelligence Index chart showing Astra scoring 61 alongside GPT-5.6 Sol.
• [06:37 - 07:21] CodeRabbit Security Demo: Walkthrough of CodeRabbit Security scanning codebases, showing reachability and blast radius assessments, and automated pull-request remediation.

Claims & numbers

• Anthropic Claude Fable 5.1 / Mythos 5.1:
  • The presenter states Fable 5.1 solved a 5-year-old, 1-in-a-million crash for hedge fund Millennium by disassembling vendor binary code into raw assembly.
  • The presenter notes Mythos 5.1 achieves ~50% hit rate in viable protein binder design across 12 targets, compared to the industry standard ~10-15%.
  • Pricing is claimed at $10 per million input tokens and $50 per million output tokens.
• Meta Muse Spark 1.3:
  • Mark Zuckerberg claimed Muse Spark 1.3 has frontier performance that is "almost too cheap to meter."
  • Standard API pricing is stated as $1.25 input / $4.25 output per 1M tokens ($0.15 cached input); Contributor tier is $0.10 input / $0.20 output ($0.002 cached input) in exchange for Meta training on prompt data.
  • Alexandr Wang claimed a "meaningful double digit" percentage of developers are selecting the contributor tier.
• OpenAI GPT-6 Astra:
  • Greg Brockman declared: "welcome to the AGI era."
  • OpenAI claims Astra was pre-trained on 100,000+ GPUs at the Stargate site in Texas, with previous models providing a significant portion of supervision during training.
  • The presenter states Sam Altman confirmed the model underwent a formal review process with the Trump administration prior to release.
  • On OSWorld 2.0, Astra reportedly scored 73% in ~40 minutes per task (compared to GPT-5.6 Sol at 65% in 75 minutes).
  • OpenAI reported Astra scored 100% on ExploitBench, 42.4% on Exploit Gym, 64.6% on Terminal-Bench Science 0.1, 97.6% on FrontierMath Tier 4 (v2), and 99.9% on ARC-AGI-3 using OpenAI's response API harness (66% on the standard harness).
  • OpenAI reported Astra meets the "Critical" cybersecurity threshold under its Preparedness Framework, enabling autonomous end-to-end zero-day discovery and exploitation.
  • Standard API pricing is stated at $10 per million input tokens and $50 per million output tokens (with Fast mode offering up to 2x speed at 2x price).
  • In Artificial Analysis's independent Intelligence Index, Astra scored 61 (matching GPT-5.6 Sol, but trailing Claude Fable 5.1's score of 66).

Notable quotes

• [00:03] "The years start coming and they don't stop coming, and I've never understood those words more deeply than I did this week after what felt like years in AI bizarro land."
• [04:10] "Sam told him something we've all heard after making a desperate late-night plea: go to bed."
• [06:33] "It scored a 61, which is exactly the same as GPT-5.6 Sol, and 5 points behind Claude Fable 5.1. Something doesn't add up here..."

Assessment
This is a tech news roundup and critical commentary video reviewing the launch events and early documentation of Claude Fable/Mythos 5.1, Muse Spark 1.3, and GPT-6 Astra. The presenter did not have hands-on early access to GPT-6 Astra himself, relying instead on official benchmark releases, public social media demonstrations by authorized early testers, and independent evaluation data from Artificial Analysis.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.