The day's most important AI news: breakthroughs, releases, funding, and policy — curated for developers, founders, and investors.
Today's Stories
Claude Opus 4.8 ships with adaptive thinking, 1M-token context, and 3x cheaper Fast mode
Vendor: Anthropic
Anthropic released Claude Opus 4.8, its upgraded flagship featuring dynamic reasoning-effort adjustment, a 1M-token context window, and improved coding/knowledge work performance at the same headline price ($5/M input, $25/M output). Fast mode is now 3x cheaper and 2.5x faster. The launch dominated Hacker News with 1729 points and 1346 comments, and is already available on AWS Bedrock, Perplexity Max, Kiro IDE, and Vertex.
Anthropic raises $65B Series H at $965B valuation, briefly topping OpenAI
Vendor: Anthropic
Anthropic closed a $65B Series H at a $965B post-money valuation, briefly overtaking OpenAI as the most valuable private AI lab. The round coincides with the Opus 4.8 launch and a reported $47B ARR, though analysts question whether the revenue is inference-inflated rather than durable.
AWS Bedrock launches AgentCore Payments — AI agents can transact via x402 micropayments
Vendor: AWS
AWS introduced AgentCore Payments in Amazon Bedrock, letting AI agents autonomously discover, evaluate, and pay for web content, APIs, and services via the x402 protocol. The preview was built with Coinbase (wallet infrastructure) and Stripe (payment rails), aiming to make agents 'economically active' within a single execution loop.
Microsoft builds Copilot 'super app' and ships in-house coding model at Build, distancing from OpenAI
Vendor: Azure
Fortune reports Microsoft is building a Copilot 'super app' that unifies coding, chat, and AI tools into a single product, while The Information reports a new in-house Microsoft coding model debuts next week at Build. MSFT rose ~3% on the coding-model news as the company visibly distances itself from OpenAI, Anthropic, and Google dependencies.
ChatGPT Codex brings Computer Use to Windows, plus remote control from iOS/Android
Vendor: OpenAI
OpenAI extended Computer Use in the ChatGPT Codex app to Windows, letting eligible users have ChatGPT interact directly with Windows applications for testing, debugging, and refining. Users can also remote-control their Windows machine via ChatGPT mobile or Codex on Mac. The update adds Codex Profiles for tracking activity and token usage.
Google DeepMind ships Gemini Embedding 2 — native multimodal RAG across text, image, video, audio, code
Vendor: Google
Google DeepMind released Gemini Embedding 2, a native multimodal embedding model available via the Gemini API and Vertex AI. It supports retrieval across text, image, video, audio, documents, and code, and Google claims state-of-the-art performance across multiple embedding benchmarks — aimed squarely at advanced RAG and multimodal search.
Mistral launches Vibe consumer agent and Industrial Engineering stack on new 128B Medium 3.5
Vendor: Mistral
Mistral AI rebranded its consumer assistant to 'Mistral Vibe' (with Work Mode for long tasks and Code Mode for remote coding) and launched 'Mistral for Industrial Engineering,' combining LLMs with Emmi AI physics simulation. Both run on the new flagship Mistral Medium 3.5 — a 128B dense model that self-hosts on as few as 4 GPUs. Mistral also announced a 10MW inference data center in Les Ulis and Microsoft Copilot Studio integration.
Alibaba Qwen3.7-Max ranks #4 globally on Code Arena WebDev, beating OpenAI and Google
Vendor: Alibaba
Alibaba's Qwen3.7-Max reached #4 on Code Arena's WebDev leaderboard, the only non-US developer in the top five (behind several Claude models) and beating deployed OpenAI and Google models on web application building. Alibaba also reported Qwen3.5 hitting 580 tps on the TokenSpeed engine.
Apple readies dramatic Siri overhaul for WWDC 2026, reportedly powered by Google Gemini
Vendor: Apple
Apple is preparing a dramatic Siri overhaul for WWDC 2026, transforming it into a capable AI agent for multi-step tasks. The new Siri will reportedly be powered by Google Gemini (with third-party AI options) and include a dedicated app, Dynamic Island integration, and AI photo editing. Bloomberg separately reports OpenAI is weighing legal action against Apple over the stalled ChatGPT-Siri integration.
xAI ships grok-build-0.1 coding API and pledges Grok 4.2 open-source by year-end
Vendor: xAI
xAI launched grok-build-0.1 in public beta via the xAI API at $1/M input and $2/M output — the model powering the Grok Build CLI agentic coding tool — and launched a daily public build-tracking page. Elon Musk separately pledged to open-source the 0.5T-parameter Grok 4.2 base model by end of 2026, with Grok 5 training on the 500,000-GPU Colossus 2 cluster.
NVIDIA raises Taiwan spend to $150B/year and begins first Vera CPU deliveries
Vendor: NVIDIA
Jensen Huang said NVIDIA is raising annual Taiwan manufacturing and supply-chain spend from $100B to $150B to meet AI demand, with the first Vera CPUs shipping to leading AI labs and clouds. Huang floated a path to $3T in revenue driven by agentic inference. NVIDIA AI's X account also teased 'a new era of PC.'
Malicious npm package 'mouse5212-super-formatter' targets Claude users via supply-chain attack
Vendor: Anthropic
A malicious npm package named 'mouse5212-super-formatter' was found exfiltrating files from the Claude AI user directory to an attacker-controlled GitHub repo, disguised as an archive utility and executing during postinstall. The attack underscores supply-chain risk in AI dev workflows as Claude Code adoption surges.
Robinhood opens platform to AI agents trading stocks
Vendor: Other
Robinhood opened its platform to AI agents trading stocks on behalf of users, marking one of the first production deployments of agentic finance at a major retail broker. The move raised fiduciary, liability, and governance questions in a Hacker News thread that drew 91 points and 166 comments.
Hugging Face open-sources $2,500 3D-printable humanoid robot
Vendor: Hugging Face
Hugging Face open-sourced complete plans for a bipedal humanoid robot platform costing roughly $2,500 in parts. The design uses 75 3D-printed files, common actuators and electronics, and a public Onshape CAD model for inspection and modification — extending the LeRobot project into humanoid form factor.
Liquid AI ships LFM2.5-8B-A1B MoE trained on 38T tokens
Vendor: Other
Liquid AI revealed an 8B-A1B mixture-of-experts model trained on 38 trillion tokens, intensifying the small-efficient-model race. The release drew 159 points on Hacker News as another small-MoE entrant challenging dense incumbents.