Jump within the current audio brief. Story thumbnails mirror the cards below when available.
Same-day briefs for Thu Oct 8, sorted by release time. The selected date stays fixed.
Changing the date preserves the current brief type: AI News AM.
/ai-news/2026-10-08/.Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
OpenAI published a public repository of 722 manuscripts organized into 372 result families, produced by an internal frontier model run against roughly 4,000 math problems at about three hours of compute each. Many ship with computer-checkable Lean formalizations, with more promised. The math community is still verifying, and skepticism is warranted given earlier OpenAI math claims that got walked back.
Haiku 5.5 lands at ten cents per million input tokens and fifty cents output for prompts up to 100K, roughly 75% below Haiku 4.5 and dead level with GPT-6 Luna. Computer use is now in the small model. The cheap tier is where the actual volume war is being fought.
Pro, Plus, Business and Enterprise users now get GPT-6 in the Chat tab immediately. The bigger shift is Intelligent UI — interactive charts, buttons and editable diagrams rendered inline instead of walls of text. ChatGPT is quietly becoming an application surface, not a text box.
Mistral's biggest model ever: a granular MoE with 1.05T total and 49B active parameters per token, a 1.6B-parameter vision encoder, and a one-million-token context window. Preview is live via Mistral Studio and API at $1.36 in / $4.18 out per million tokens. Open weights are slated for roughly October 31.
Common Sense Media's Youth AI Safety Institute ran 4,000-plus prompts and found ChatGPT missed more than one in four warranted crisis referrals, let self-harm and disordered-eating conversations run up to an hour without alerting a parent, and let teens bypass study-mode limits by deleting an @study prefix. OpenAI disputes the methodology, saying linked teen and parent accounts weren't fully configured during testing.
Boyu Capital and IDG Capital co-led, with Tencent, HSG and ZhenFund returning. Bloomberg puts the valuation at roughly $4 billion — double the price Meta had agreed to pay before Beijing killed the $2 billion acquisition. Manus resumed independent operations in September.
A 740M-parameter Apache 2.0 embedding model covering text, code, images, video and audio, with a reported 9.92-point gain on MTEB Code. Modular loading scales it down to 270M parameters for on-device retrieval. Boring on the surface, load-bearing for every RAG stack underneath.
The GPU cloud is lining up a mega-round ahead of a planned 2027 IPO. The eye-popping number is the backlog: roughly $15 billion in June to about $50 billion by September. Reported, not confirmed by Lambda.
JFrog disclosed CVE-2026-105192, a CVSS 9.8 remote-code-execution flaw in LMCache through version 0.5.5 — a caching layer sitting inside many vLLM deployments — with no patch available. Same window, Pwn2Own Ireland paid out $388,500 on day one for 32 zero-days, including a full-takeover chain on LiteLLM. The inference supply chain is the soft underbelly.
September came in at NT$511.86 billion, with Q3 hitting NT$1.494 trillion — up 51% year over year. Management credits surging AI accelerator demand from the major chip designers. Whatever else you believe about the bubble, the foundry numbers are real and still accelerating.
