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

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

More

Summary

Fireship’s Code Report for September 4th, 2026 covers one of the most consequential weeks in AI history in its signature rapid-fire style. The video opens with Anthropic’s release of Fable and Mythos 5.1, highlighting two standout case studies: Millennium’s hedge fund using Fable to reverse-engineer a vendor binary and locate a years-old crash bug, and Mythos 5.1 raising protein design success rates from 10% to 50%. Meta’s MuSpark 1.3 arrives next, noted for its aggressive pricing — $1.25 per million input tokens standard, $0.10 with a contributor data-sharing agreement — with Alexander Wang reporting double-digit developer uptake on the cheaper tier.

The centerpiece is OpenAI’s GPT-6 Astra, framed as a chaotic launch: the announcement page briefly went dark, embargoed press stories ran early, and ChatGPT, Claude, Grok, and Cursor all went offline simultaneously around the same time (likely an Azure outage). Astra itself is described as trained on over 100,000 GPUs at the Stargate Texas facility, with the first significant contribution from prior models during training. Key benchmark results include OSWorld computer-use at 73% versus Soul’s 65%, 99% on ARC-AGI 3, 100% on Exploit Bench, and a preparedness-framework designation as the first model to hit the critical cyber threshold.

Pricing lands at $10 per million input tokens and $50 out — the same as Fable 5.1 — and rollout to Plus and Pro subscribers was pending at time of recording.


📺 Source: Fireship · Published September 04, 2026
🏷️ Format: Roundup

1 Item

Channels

3 Items

Companies