Wednesday, Nov 18, 2026

AI;DR
2026-11-18 · 14 stories- openaiOpenAI’s GPT-6.2 Astra became the first model to score 100% on ARC-AGI-3, but early testers called the release frustrating and a step back in writing and coding.+1
- nvidiaNew US rules block foreign data centers from remotely serving advanced AI chips to Chinese customers, slowing Tencent’s next model. Nvidia also stopped selling AI chips to China, while Southeast Asian hubs brace for fallout.+2
- reasoningCRAM, a learned classical energy model, solved 99.996% of Sudoku-Extreme and 100% of Maze-Unique at least 1,000× faster than looped models, using far fewer inference FLOPs.
- anthropicAnthropic is selling $26B in stock as Nvidia and Google each invest $10B; AMD adds $2.5B, while Microsoft, Broadcom, Micron and TSMC also participate. The IPO will trade on Nasdaq as ANT.
- agentA Harvard study found AI coding agents increased code output 30%, but finished work did not measurably rise and code review took 49% longer, shifting the bottleneck to review.
- microsoftMalicious versions 1.26.1 and 1.26.2 of Microsoft’s agent-framework package stole Azure identities, GitHub OIDC tokens and CI/CD secrets. PyPI removed them; users should pin 1.26.0 and rotate credentials.
- nvidiaNvidia stopped reporting its Singapore unit’s revenue in 2025 after it approached half of global revenue; Financial Times says the unit rose 72% year over year to $104B, amid a 70% surge in revenue booked there.
- openaiOpenAI is giving Apollo Research, METR and two other independent auditors access to unreleased models, source code and training data for the first time. OpenAI can redact or delay reports, limiting how independent oversight is.
- openaiOpenAI changed GPT-6.2 Astra and Sol extensively in the last 3–4 days before launch without early reviewers retesting them. Epoch found Astra can perform 12.3 serial reasoning steps per output token, complicating monitoring.
- agentKPMG used more than 100 AI agents to run a parallel audit of a FTSE 250 company, saying they asked better questions and dug deeper than staff. Human intervention was still needed.
- inferenceHugging Face now routes language-model requests through OpenRouter, which chooses the cheapest or fastest provider and cut capacity errors 46%. Bring-your-own API keys are going away.
- openaiByteDance is in advanced talks with Microsoft to spend about $500M more on Azure, mainly to access OpenAI models. The plan has drawn complaints from OpenAI.
- agentTeleChat’s Xing4 advances frontier agentic models on Ascend chips and MindSpore, including a 285B-parameter model trained on 8,192 Ascend 910C cards and the TeleGym agent platform.
- data-centerA Harvard study in Nature Sustainability found 403 US hyperscale data centers emitted 25.7 million metric tons of CO2 from May 2024 to April 2025, with over half their electricity from fossil fuels.