Anthropic launched Claude Opus 4.6 with 1M context and adaptive thinking, demonstrating orchestrated agent teams by building a C compiler. OpenAI countered with GPT-5.3-Codex, a new coding-focused model, while Google researchers detailed a sequential attention technique that slashes model size without accuracy loss. An 8B world model also beat 402B Llama 4 by generating web code instead of pixels, with open weights released.

ACE-Step-1.5, an MIT-licensed open source audio generative model, launched with performance nearing commercial platforms like Suno. French authorities raided X's offices while the UK opened a separate probe into Grok AI. Apple integrated coding agents into Xcode 26.3, and researchers disclosed a supply chain attack on Notepad++.

xAI is merging with SpaceX in a major consolidation move. Anthropic released Claude's constitution and published new research on how alignment scales with model intelligence, while Kimi K2.5 Thinking became the top open-weights model on the Extended NYT Connections benchmark.

Mistral released Vibe 2.0, an updated conversational AI model with improved capabilities. Separately, Google Research published findings on when and why AI agent systems work, and the CISA breach drew attention after the agency's acting director reportedly uploaded sensitive documents to ChatGPT.

Researchers proposed using LLMs as semantic regularizers for enumerative feature synthesis in decision trees, aiming to filter out nonsensical features while preserving interpretability. Alibaba's Qwen team indicated its next major model will prioritize quality over release speed, signaling a shift in Chinese frontier model development. Community discussion highlighted a claim that ChatGPT 5.2 Pro solved Erdős Problem 281, intensifying debate over AI mathematical reasoning.

ML hiring burnout is escalating as candidates from top labs report months of rejections and mental exhaustion, signaling structural problems in research recruiting. ClickHouse closed a $400 million Series D, acquired Langfuse, and launched a Postgres integration. DeepSeek released Engram, a static memory unit design that avoids recomputing fixed knowledge through transformer layers.

PACEvolve, a new framework for integrating large language models into evolutionary search, was published today, offering structured scaffolding to fix reliability problems in AI-driven optimization. Zhipu AI and Huawei released GLM-Image, China's first state-of-the-art multimodal model trained entirely on domestic Ascend 910 chips, while Black Forest Labs shipped FLUX.2 [klein], a 9B image generator that runs in under a second on a 4090.

Zhipu AI trained its first major model entirely on Huawei's hardware stack, breaking US chip reliance. Mistral published the Ministral 3 series paper detailing 3B, 8B, and 14B parameter models, while NVIDIA released Orchestrator-8B for routing tasks to specialized tools and models.

Researchers released APEX-SWE, a new benchmark measuring whether frontier AI models can execute economically valuable software engineering work rather than just narrow coding tasks. ZAI open-sourced GLM-Image, a hybrid autoregressive and diffusion model matching mainstream image generators while excelling at text rendering. Google also pushed into clinical AI with MedGemma 1.5, adding improved medical image interpretation and speech-to-text capabilities for diagnostic workflows.

Baichuan AI released Baichuan-M3-235B, a 235-billion-parameter medical-enhanced LLM that scales up its specialized healthcare capabilities from the previous 32B model. DeepSeek open-sourced Engram, a conditional memory framework introducing a new sparsity axis for LLM efficiency. Microsoft also released FrogBoss 32B and FrogMini 14B, coding agents fine-tuned on Claude Sonnet 4 debugging trajectories.