Good morning. The Astra rollout we’ve been tracking all week is finally here, and OpenAI is calling it the start of the “AGI era” — a phrase Greg Brockman is happy to say out loud even as his own chief scientist warns that watching what these models are actually thinking is getting harder. Meanwhile, Nvidia decided to spend nearly $13 billion on Hugging Face, and OpenAI decided a billion dollars of Cursor revenue wasn’t worth doing business with Elon Musk.
GPT-6 Astra launches with a 99.9% ARC-AGI-3 score and a lot of asterisks. OpenAI shipped Astra yesterday, with rollout starting on enterprise cybersecurity customers before expanding to paid tiers. It’s the first OpenAI model to cross the company’s “critical cybersecurity capability threshold” — a designation that reads differently given last week’s Hugging Face incident. Wired notes chief scientist Jakub Pachocki flagged that internal-reasoning monitoring is degrading as models scale, and OpenAI may slow development if visibility gets worse. TechCrunch reports the model uses “opaque recurrence,” which makes chain-of-thought auditing harder — exactly the tradeoff Pachocki is worried about.
The ARC-AGI-3 numbers are impressive but contested. ARC Prize’s own writeup shows Astra hitting 62.7% with a standard harness and 99.9% with a custom “Provider Adapter” harness, using fewer actions than the median human on 96% of levels. One HN commenter flagged that the scorecard shows GPT-5.6 Sol at 7.8% while the same page admits Sol would score around 30% with the newer harness — a comparison that flatters Astra’s jump. Others pointed to the Erdős problem benchmark, where Astra solved just 2 of 68 problems. And Artificial Analysis has Astra at 61 on intelligence, behind Anthropic’s Opus 5. The most-liked HN comment was more philosophical: is solving a snake-like puzzle in the fewest moves really what defines intelligence?
Nvidia buys Hugging Face for $12.9 billion. Jensen Huang confirmed the acquisition, first rumored in late August, at nearly 3x Hugging Face’s 2023 valuation. Both The Verge and Wired emphasize the promise to keep the platform cloud- and hardware-agnostic. CNBC reports that HF CEO Clément Delangue initiated the deal himself over the summer. HN was split: some celebrated the founders’ payday, others compared it to buying Docker Hub in 2018, and several noted the conflict of interest given Nvidia’s heavy investment in closed-model labs. One commenter had the punchline of the day: can Nvidia now direct Hugging Face to sue OpenAI over the hack?
OpenAI walks away from a billion in Cursor revenue to avoid Musk. After SpaceX’s $60 billion acquisition of Cursor, OpenAI is ending the partnership, citing an inability to trust Musk-owned companies to honor its terms of service. Cursor was a top-five OpenAI customer projected to clear $1 billion in annualized revenue. Cursor’s CEO countered that OpenAI models handle only about 5% of Cursor traffic; OpenAI replied that token counts don’t map to revenue. The feud is now costing real money on both sides.
Cerebras hosts Qwen 3.8 27B at 1,500 tokens/sec — if you can stay under the rate limit. The model is live on Cerebras with 128k context and unpruned weights, quantized only for storage. HN reactions ran hot and cold: the output speed is genuinely wild, but the 150k–450k TPM caps are a real constraint. One commenter hit the limit in 90 seconds and burned through $1.10, partly because cached tokens count against the quota. Multiple users are asking for OpenRouter availability so they can actually use it for sustained coding sessions.
That’s the day: a new frontier model that both sets records and quietly makes safety monitoring harder, a $13 billion open-source acquisition, and a reminder that in 2026 corporate grudges can outweigh billion-dollar contracts. More tomorrow.