Thursday, Nov 19, 2026

AI;DR
2026-11-19 · 15 stories- googleGoogle deployed Gemini 4 Neon to AI Studio at $3 per 1M input tokens and $15 per 1M output tokens. Source code also references potential model Cobalt, suggesting further Gemini 4 expansion.
- llmSRE-1 is a 27B model built for production incident investigations. Its creators say it leads the price-performance frontier for this task.
- KPMG used more than 100 AI agents to audit a FTSE 250 company in parallel with staff during a pilot. The agents asked “better questions” and dug deeper, but humans still had to intervene.
- AI safety, open-source, science, law and e/acc advocates signed the MisAnthropic Statement before Anthropic’s planned IPO. Dan Hendrycks said some safety groups declined to sign over funding and relationship concerns.
- anthropicAnthropic raised $81.31B in an upsized IPO at $83 a share, plus a $26B private placement, taking proceeds above $107B. The deal surpassed SpaceX’s $86.25B listing and valued Anthropic at $1.08T.
- llmResearchers trained a black-box inverse language model from scratch to reconstruct prompts from LLM responses without weights or logits. It transfers across datasets and models and beats prior work on token-based metrics.
- googleGoogle is rolling out Project Genie with the Genie 4 world model to eligible Google AI Ultra subscribers. The release opens access to a new world-model experience.
- regulationCalifornia Gov. Gavin Newsom proposed broad state powers to regulate AI and impose strict safety measures. The plan stops short of requiring companies to develop kill switches.
- fundingThe Homeland Security Task Force launched the Threat Disruption Center to fight crypto fraud and cybercrime, using TRM Labs’ AI platform through a roughly $95M no-bid deal. The effort targets rising attacks on Americans.
- googleA test of ChatGPT, Gemini, Google AI Overviews and AI Mode asked the same 210 buying questions on two consecutive days. 84% of answers named different brands, while smaller brands lost half their visibility.
- CRAM, an energy-based reasoning model, solved 99.996% of Sudoku-Extreme and 100% of Maze-Unique after 42ms and 132ms of training on one GPU. It was at least 1,000× faster than looped models.
- New US rules bar foreign data centers from remotely serving advanced AI chips to Chinese customers; Nvidia also restricted sales to providers renting over 50% of GPUs to China. Tencent said its next model will slow.+1
- Tencent released open-weight Hy4 Flash, a 224B MoE model with 12B active parameters and a 1M-token context window. It slightly beats 770B Hy4 Preview on productivity tasks at one-thirteenth the cost.
- llmWorkers AI’s new Superintelligence team released updated Clairic and Clairic-flash as open-weight extraction models. Clairic has graduated from preview, while Clairic-flash joins the Workers AI lineup.
- openaiPalo Alto Networks agreed to buy Applied Compute for $4.5B in stock, up from its $3.25B valuation earlier this month. Founded last year by former OpenAI researchers, the startup customizes open-weight AI models.