Anthropic filed an amended S-1 showing nearly $17B in Q3 revenue but a $567M operating loss driven by compute costs, with about $895B in infrastructure commitments. Meituan launched LongCat-2.5 (1.6T parameters) for long-horizon coding agents, while SemiAnalysis's new GDPx benchmark finds the best model scores only 12% on expert-level economic tasks. On adoption, Abu Dhabi signed a set of deals with OpenAI, and Goldman Sachs narrowed Claude Code and GitHub Copilot access to engineers only.
The amended IPO filing submitted Nov 2 shows revenue of nearly $17B for the quarter ended Sept 30 (up more than $5B from Q2, roughly 12x year over year), but the company swung to a $567M operating loss, mainly due to $5B of compute under short-term commitments. Total compute and infrastructure commitments are about $895B; Anthropic has not disclosed a price range, share count or IPO date.
Following a meeting between Abu Dhabi's Crown Prince and Sam Altman, the two sides launched a 'Machine-Readable Government' initiative that converts regulations into machine-readable form so AI can help check compliance (first pilot: food imports), processed on GPUs in the UAE. Mubadala, Inception42 and OpenAI also agreed to use AI for investment analysis, plus a five-year partnership to bring frontier AI to UAE government, enterprise and education; contract values were not disclosed.
Per Bloomberg (Nov 2), Goldman's October internal guidance limits AI coding tools to about 12,000 engineers, removing access for bankers and traders who had used them for months to monitor trading flows and produce materials. The bank cited a balance of productivity and risk management — a sign financial firms are reining in broad rollout of coding agents.
The 1.6-trillion-parameter multimodal model (about 48B active per token) was further trained on long task sequences and large-scale reinforcement learning, turning real agent failures into reproducible training tasks. LongCat-2.5 is available via Meituan's chat service and API; GitHub and Hugging Face releases are planned.
GDPx measures agentic capability across 11 high-value domains (chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement and more). Claude Opus 5.5 leads overall, GPT-6 Astra leads in self-improvement and materials science, and Gemini 4 Carbon in biology and software engineering; finance and legal tasks are so hard that nearly every model scores 0 under binary grading.
Omar Khattab's group describes 'Weight-Space Programming', where a frontier model reads and edits weights from a natural-language spec — e.g. editing 0.016% of Qwen3.5-0.8B's weights to stop reasoning loops, lifting MATH-500 from 7.9% to 48.1%. This is a blog post, not yet peer-reviewed.
DeepSeek's open-source agent tool now lets users describe an idea and get a ready-to-use plugin, with examples such as a design workspace, a sandbox game and a 3D LLM visualizer. It follows last week's Super-agent design and signals DeepSeek's push into agent products; Harness remains in preview.
President Lee Jae Myung presented a 2027 budget with 21.3 trillion won (about $16B) for semiconductors, data centers, physical AI and AI, plus 4.6 trillion won to acquire a world-leading sovereign AI model — the race for national AI in Asia keeps accelerating.
If these digests help you make sense of technology, a small contribution keeps NHỊP picking out what is worth reading.
Scan the QR code with a Vietnamese banking app.
Download QR ↓ · View full image ↗