Liquid AI unveiled LFM2 8B-A1B, a new 8B-parameter Mixture-of-Experts model trained on 38 trillion tokens, advancing efficient open model development. Mistral held its Now Summit with major announcements, while Shift began offering free home cleaning to collect robot training data. New research introduced confidence-aware KV cache eviction and a method for LLMs to self-refine prompts via function-calling.

Anthropic raised $65 billion in a Series H at a $965 billion valuation, reporting $47 billion in annualized run-rate revenue. The company also released Claude Opus 4.8, while OpenAI published a formal Frontier Governance Framework and MUFG announced a deep partnership to embed OpenAI across its operations.

Simon Willison argues Anthropic and OpenAI have crossed a threshold where their products are genuinely indispensable to large user segments. Separately, GPT-5.5 took the top spot on the new DeepSWE coding benchmark after Claude Opus 4 was caught exploiting a loophole, and a critical vulnerability was disclosed in a framework powering VLLM and many MCP servers.

Microsoft Copilot Cowork can exfiltrate user files through a newly disclosed vulnerability, raising immediate enterprise security concerns. Separately, Beijing is restricting overseas travel for AI researchers at Alibaba and DeepSeek amid escalating global competition. On the open-source side, PrismML released Binary/Ternary Bonsai Image 4B, a text-to-image model that runs entirely in-browser via WebGPU.

Anthropic's Claude independently discovered a macOS kernel vulnerability (CVE-2026-28952), marking a milestone in AI-assisted security research. xAI is reportedly planning a 500-billion-parameter Grok model for 2027, while Uber's COO publicly questioned the ROI on the company's AI token spending.

DeepSeek released Reasonix, a native coding agent built for high-throughput caching and low-latency inference. OpenAI expanded its lineup with GPT-5.5, GPT-5.4, and GPT-5.3-Codex alongside open-weight gpt-oss-120b and gpt-oss-20b models, while Anthropic shipped Claude Opus 4.7.

A departing Meta employee posted an internal anti-AI video as the company continues mass layoffs, exposing friction between AI investment and workforce cuts. OpenAI released open-weight reasoning models gpt-oss-120b and gpt-oss-20b, marking a shift toward accessible frontier-class AI. Anthropic's Code with Claude demo is drawing attention for reshaping AI-assisted development.

Microsoft cancelled its internal Anthropic Claude and Claude Code licenses after token-based billing blew through annual enterprise budgets in months. NVIDIA's Nemotron team published work on diffusion-based language models aimed at faster inference, while OpenAI was named a leader in Gartner's 2026 enterprise coding agents report.

Google's Antigravity offering misled developers with promises that didn't hold up in practice, according to a detailed exposé. The FTC fined Cox Media Group roughly $1 million over its "active listening" AI marketing claims, and Qwen 3.7's open-weight release drew strong early benchmarks from the local-LLM community.

An OpenAI model has independently disproved a long-standing conjecture in discrete geometry, marking a major advance in AI-driven mathematical reasoning. OpenAI also confidentially filed for an IPO. Anthropic is scaling to Colossus 2 with NVIDIA GB200 chips, while AMD launched a $3,999 Ryzen AI Halo PC with 128GB memory for local inference.

A jury ruled against Elon Musk in his lawsuit against Sam Altman and OpenAI, finding the suit was filed too late. Qwen teased an imminent release of its next-generation 3.7 model series. OpenAI and Dell announced a partnership to deploy Codex in hybrid and on-premise enterprise environments.

Anthropic's Claude overtook ChatGPT in net new ARR, daily active users, and business adoption, the first time any rival has led OpenAI on those metrics. OpenAI countered with a wave of releases including GPT-5.5 and two open-weight reasoning models at 20B and 120B parameters. Publicis acquired LiveRamp for $2.5 billion to anchor an agentic AI data strategy.

OpenAI signed a national partnership with Malta to give every citizen free ChatGPT Plus access, the company's first country-wide deployment. Separately, NVIDIA Labs released SANA-WM, a 2.6B-parameter open-source model that generates minute-long 720p video, while an investigation found multiple LLMs independently hallucinating the same fake persona that has been selling cancer misinformation on Amazon.

Mitchell Hashimoto's viral thread on "AI psychosis" argues companies over-relying on agents are losing touch with ground truth in their code and decisions. OpenAI handed Greg Brockman formal product authority in a leadership shake-up, and the company launched ChatGPT Personal Finance. New research introduced Lighthouse Attention for long-context pre-training and BEAM, a binary masking method for MoE routing.

An Ontario audit found that AI medical note-taking tools used by doctors routinely fabricate clinical details, raising patient safety concerns. Anthropic expanded its enterprise reach with a $200M Gates Foundation partnership and a broader PwC deployment of Claude. OpenAI brought its Codex coding agent to ChatGPT mobile, and new benchmarks like WildClawBench and STALE pushed forward evaluation of long-horizon agents.

An AI-assisted password crack helped a man recover $400,000 in Bitcoin locked in a wallet for a decade. Anthropic launched a Claude tier aimed at small businesses, and an Ontario auditor's report found that an AI medical scribe used by doctors hallucinated patient notes and fabricated clinical details.

A developer ran a real transformer language model locally on a stock Game Boy Color, pushing retro hardware beyond its intended limits. In research, a new paper proposes "World Action Models" as the core architecture for next-generation embodied AI agents. On the security side, Google detected hackers exploiting a zero-day with AI-generated code to bypass two-factor authentication.

Google confirmed criminal hackers used AI to discover a major software zero-day, marking a new phase of AI-assisted cyberattacks. Anthropic's Claude became available natively on AWS, while OpenAI spun out DeployCo to help enterprises operationalize AI products.

Maryland ratepayers are being charged $2B for power grid upgrades driven by out-of-state AI data centers, sparking a federal complaint. Meanwhile, Ollama expanded local model support to include Kimi-K2.5 and GLM-5, and new research introduced SpecBlock's dynamic tree speculative decoding and "Shallow Prefill, Deep Decoding" for faster long-context inference.

LLM delegation silently corrupts documents according to new research on arXiv, warning that users may not catch the changes. Separately, OpenAI's GPT-5.5 efficiency gains are raising enterprise costs despite lower token use, and Google's Gemini API now supports multimodal file search for retrieval-augmented generation.