OpenAI ships GPT-6 Astra — and calls it the AGI turning point
OpenAI released GPT-6 Astra this week as a limited preview. Daybreak enterprise partners get it first, followed by ChatGPT Plus/Pro/Business/Enterprise, then the official API and AWS.
- Reported scores: 99.9% on ARC-AGI-3 (no harness) and 100% on ExploitBench, including a post-cutoff refresh built from freshly disclosed vulnerabilities.
- It produced original math: a Lean-verified proof that prime gaps of at most 186 recur infinitely, and 2 of 68 open Erdős problems solved on FrontierMath. Epoch AI logged a new Effective Compute Index of 169.
- It is OpenAI's first model with a "Critical" cybersecurity designation — initial rollout is restricted to vetted partners with White House coordination. A recurrent-depth technique drives the capability but obscures intermediate reasoning, reviving the "can we audit its reasoning?" debate under EU AI Act transparency rules.
- Greg Brockman: "welcome to the AGI era." The institutional tension isn't the label — it's a model that aces exploit benchmarks while reducing visibility into how it thinks.
Nvidia buys Hugging Face for $12.9B
Nvidia confirmed a $12.93 billion deal to acquire Hugging Face — the hub hosting ~3M models, ~1M applications, and ~500k datasets, used by ~18M developers.
- Jensen Huang pledged Hugging Face stays open and that "Nvidia compute will not be required." The deal is subject to regulatory approval and expected to close in H1 2027.
- The company that controls GPU supply now also owns the primary distribution channel for open-weight models. For any team whose AI strategy leans on HF neutrality, the question is whether the platform stays genuinely vendor-neutral or gradually tilts toward Nvidia infrastructure. Huang's openness pledge will be tested by commercial gravity.
Four frontier labs went down at once
On September 3, Anthropic (Claude), xAI (Grok), OpenAI (ChatGPT/Codex), and Google (Gemini) all suffered overlapping outages — roughly 3 hours 40 minutes of degraded service, with 37,000+ OpenAI incident reports at the peak.
- Root cause pointed to a routing error at OpenAI and an infrastructure issue at Anthropic; Cursor went down on the upstream dependency.
- The rare synchronization is a reminder that top-tier LLMs share fragile cloud dependencies. When the whole frontier tier blinks together, "which model is best" matters a lot less than "is the substrate up?" Multi-vendor strategies look less like diversification when everyone rents the same glass.
More daily AI briefings at https://ainexusdaily.vercel.app
Top comments (0)