Daily Highlights
- OpenAI wraps DevDay 2026 with 20+ major announcements
- GPT-6.2-Astra enters gray-scale testing
- Claude Mythos 5.5 briefly spotted on Anthropic site
- Anthropic posts Q3 revenue of $17B, back in the red
- Apple-Qwen integration opens developer Beta on Nov 4
Today's Headlines
[OpenAI wraps DevDay 2026: 20+ launches as ChatGPT opens as a shared agent platform]On November 3, OpenAI published its DevDay 2026 recap - its biggest yet, with more than 20 major announcements across ChatGPT, Codex and its model lineup. The core theme is agents taking on ongoing responsibilities: ChatGPT is now open as a shared surface where humans and agents collaborate, and developers can launch native experiences directly to 1.2B weekly users. The week also brought a string of hires: Jill Hazelbaker joins as Chief Corporate Affairs Officer and former Army Secretary Dan Driscoll as Senior Counselor for National Security Modernization. (Source: OpenAI blog)
[Anthropic's amended filing shows $17B Q3 revenue - and a return to losses]On November 2, Anthropic amended its IPO prospectus with Q3 figures: revenue of roughly $17 billion, up 47% quarter-over-quarter and more than 10x year-over-year, yet an operating loss of $567 million - profitability lasted just one quarter. Demand exceeded expectations, prompting an extra $5B in short-term compute purchases that pushed quarterly compute and infrastructure spend up 77% to $9.9B. The filing disclosed total committed spend of about $895.4 billion, nearly all for compute and data centers with payments stretching to 2045; $69.8 billion falls due within 12 months, already exceeding annual revenue. (Source: LLMHome)
[Year-end flagship race starts early: GPT-6.2 in gray testing, Claude Mythos 5.5 leaks]On November 2, OpenAI's GPT-6.2-Astra entered gray-scale testing, appearing in the model picker of some $500 Pro-tier users with routing rates from 20% to 100%. The same day, Claude Mythos 5.5 briefly appeared in the front-end code of Anthropic's website before being removed hours later; it is said to include biology research and cybersecurity capabilities, and a Hugging Face repo claiming leaked Mythos weights has topped 5,000 downloads. Both labs exposed new models in the same week, signaling a visibly shortened flagship cycle. (Source: LLMHome)
New Products & Tech
[Tokens hit e-commerce shelves for the first time as AI compute goes retail]On November 3, tokens were listed as standalone products on e-commerce platforms for the first time during China's Double 11 shopping festival. Buyers get directly consumable model calls rather than a subscription - essentially paying for compute. Metering, pricing transparency and after-sales standards are still missing; if they mature, token retailing could push AI adoption from subscriptions to pay-as-you-go, though buyers should watch for silent model downgrades and quota expiry behind low prices. (Source: LLMHome)
[ChatGPT Classic to shut down on January 30, 2027]On November 2, OpenAI notified users that the old desktop app ChatGPT Classic will stop service on January 30, 2027, with chat history migrating to the new app launched in July (Chat, Work and Codex in one). Chat and Work merge testing opens to small user groups in coming weeks, and ChatGPT and Codex usage will move to a shared quota. (Source: LLMHome)
[UBTech launches Thinker-Sim embodied-AI simulation platform]On November 2, UBTech released Thinker-Sim, an embodied intelligence simulation platform built on its Thinker-GS high-fidelity scene reconstruction algorithm. It covers real-scene digitization, large-scale synthetic simulation data and robot capability evaluation, with image-text-driven 3D asset generation, automatic collider setup, standardized task templates, live simulation streaming and task monitoring. (Source: LLMHome)
IPO & Fundings
[DeepSeek reportedly closes ~100B yuan round at ~500B yuan pre-money valuation]According to market reports, DeepSeek has completed a financing round of about 100 billion yuan at a pre-money valuation of roughly 500 billion yuan, with CITIC Securities preparing a STAR Market IPO; the parties have not responded. On October 31 its Harness team published a Super-agent architecture post describing a Router/Worker two-layer design where agents self-initiate and close the loop with reviewable PRs - about 300 messages and 20 requests per session, with p95 latency of 15.7 minutes. (Source: LLMHome)
[OpenAI trims round target to $20B, valuation cut to $1.25T]Amid an AI stock selloff and Anthropic's upcoming IPO diverting capital, OpenAI has cut its fundraising target from $30 billion to $20 billion and its valuation target from $1.4 trillion to $1.25 trillion. The Nasdaq last week logged a 2.86% single-day drop, its worst since June, as valuation pressure spreads from public markets into private AI rounds. (Source: Sina Finance)
Open Source & Research
[LeCun's new paper: a 109M-parameter JEPA beats 72B models on representation alignment]On October 29, Yann LeCun and Brown University assistant professor Randall Balestriero published a 37-page paper giving a mathematical framework for cross-modal representation alignment. Experiments show a 109M-parameter text JEPA outperforming Qwen2.5-72B and Llama-3-70B on alignment tests - a roughly 660x parameter gap - while another matched next-token prediction models trained on 900M samples using just 21.6M. The authors do not deny the value of scale, only that understanding the world need not grow monotonically with parameters. (Source: LLMHome)
Company & People
[OpenAI stacks its bench: Hazelbaker on corporate affairs, Driscoll on national security]On November 2-3, OpenAI announced Jill Hazelbaker as Chief Corporate Affairs Officer, former Army Secretary Dan Driscoll as Senior Counselor for National Security Modernization, and the new Grid Protectors program for power-grid defense. The rapid run of hires strengthens OpenAI's government-relations and safety team ahead of a potential 2027 IPO. (Source: OpenAI blog)
Policies & Data
[Google, Anthropic, OpenAI and Microsoft back a new AI crawler opt-out standard]On October 30, Google, Microsoft, Anthropic and OpenAI announced support for an upcoming website crawling opt-out standard that lets publishers declare with one switch that their content is off-limits for AI training. If it lands, it would be the first crawling-negotiation mechanism collectively endorsed by the major AI vendors, with the potential to reshape how training data is acquired. (Source: SiliconANGLE)
Disclaimer: For reference only; exercise caution for investment or decisions.
Powered by AOYII.com
