Anthropic’s Mega-Raise; OpenAI’s Statsig Play; Modeling the World
The Skim
- Anthropic closes a “historic” $13B round to scale Claude, code, and enterprise go-to-market.
- OpenAI is acquiring Statsig; its founder Vijaye Raji becomes CTO of Applications. (Terms undisclosed; Rundown pegs it around $1.1B.)
- World models jump to interactive: Tencent’s HunyuanWorld-Voyager releases code/weights; DeepMind details Genie 3.
- GovAI: Microsoft & GSA strike a deal offering Copilot to federal G5 users at no cost for up to 12 months, targeting $3B year-one savings.
- BCI “Copilot”: AI Secret spotlights a brain-computer interface assistant automating intent-level actions.
🧠 Big One: Anthropic’s $13B lightning round
What’s new: Anthropic announced a “historic” $13B capital raise ($183B valuation) to accelerate model training, productization (Claude, Claude Code), and enterprise expansion. Rundown cites ~300k enterprise customers and fast-growing Claude Code ARR.
Why it matters:
- Raises the bar on compute access and GTM for frontier labs; the capital stack signals a multi-year push to win enterprise assistant and coding markets.
- Expect faster Claude cadence (longer context, better tool use) and deeper vertical solutions as procurement cycles shorten.
Read more: The Rundown AI.
🧭 Moves, M&A and Strategy
OpenAI is acquiring Statsig; Vijaye Raji becomes CTO of Applications
OpenAI says Statsig will continue operating independently while strengthening experimentation across ChatGPT & apps; Raji will lead product engineering for Applications. Rundown pegs the deal at ~$1.1B; official post does not disclose a price.
Read more: The Rundown AI; OpenAI
Microsoft x GSA: 12 months of Copilot at no cost for G5 users
Government-wide agreement aligns with the AI Action Plan; Microsoft projects ~$3B in year-one savings via unified pricing and deployment.
Read more: The Rundown AI; Microsoft
🗺️ World Models & Generative Environments
Voyager (Tencent Hunyuan): interactive, camera-driven 3D world generation
Open-sourced RGB-and-depth video diffusion that follows a user-defined camera path and supports real-time 3D reconstruction - another step toward embodied, agent-ready environments.
Read more: The Rundown AI
DeepMind’s Genie 3
Public research post details real-time, interactive world simulation with longer consistency windows, which is useful for training generalist agents.
Read more: DeepMind
🧪 Research & Policy Round-Up
- Switzerland’s Apertus (EPFL/ETH/CSCS) — fully open LLM (8B & 70B) trained on 15T tokens spanning 1,000+ languages (≈40% non-English). Pitch: transparency (open weights + training data recipes) and sovereign deployment via Swisscom. A credible open alternative for multilingual work and public-sector use. Read more: CSCS • Swiss AI
- Mistral Le Chat “Connections” (+Memories) — Rundown calls out 20+ enterprise MCP connectors and a new Memories feature for persistent context. Official docs describe how Connections link Gmail, Calendar, Drive/SharePoint, etc., with org-grade governance—useful if you’re piloting AI in a regulated env. Read more: The Rundown AI • Mistral Help: Connections
- Microsoft x U.S. GSA — Copilot and AI services free up to 12 months for eligible federal users under a OneGov agreement; Microsoft touts $3B year-one savings and FedRAMP-aligned posture. Expect a wave of low-risk pilots across agencies. Read more: Microsoft Blog
- OpenAI for Science — Leadership shifts and program signals toward building AI platforms to accelerate discovery; expect more science-first tooling surfaced via the Apps org. (Rundown flagged early teases on social.) Read more: The Rundown AI
- Meituan’s LongCat-Flash-Chat (open-sourced) — a 560B-param MoE with ~27B activated per token, claiming strong agentic benchmarks and fast inference; MIT-licensed weights on HF. If you’re exploring agent frameworks, this is one to benchmark. Read more: Meituan Tech Blog • Hugging Face
🧪 Neuro/BCI - Toward Intent-Aware Assistants
AI-BCI “Copilot”
A brain-computer interface pipeline that translates neural activity into task-level guidance - pushing from assistive typing toward proactive, context-aware control. Early trials suggest improved speed and autonomy for users with motor or speech impairments, with broader cognitive-augmentation concepts explored. Some suggest this a path superior to Neuralink.
Read more: AI Secret
📰 Quick Hits
- Rundown highlights chip and platform moves (mobile NPUs, enterprise integrations) and media/audio updates (see tools below).
Read more: The Rundown AI.
- Agents behaving badly (and brilliantly)
AI Secret tracks LLM Werewolf/“mafia” tournaments as testbeds for deception, theory-of-mind, and coalition-forming. Reports suggest newer frontier models (e.g., GPT-5) dominate in multi-agent social reasoning. Independent research is racing to formalize these evals in multi-modal settings. Bottom line: social games are becoming stress tests for agent stacks. - Elysia cracks RAG
- NASA’s Space AI overview.
Read more: AI Secret
🚀 Tools & Launches
Gemini 2.5 Flash Image Google
What it is: A state-of-the-art image generation/editing model focused on precision edits (object swaps, recolors, background changes) with strong identity consistency and multi-image blending. Includes SynthID watermarking.
Why it matters: Tightens the gap with pro workflows; fast, controllable edits nudge AI from “toy” to “tool” for marketing, retail, and design ops. Pricing lands at ~$0.039/image via API.
Our take: Reliable characters and scenes unlock storyboard-quality pipelines and ad set variants at scale. Essentially, mock up an entire ad campaign with a set of clicks and prompts. The Rundown offers a nifty tutorial
Read more: The Rundown AI; Google Developers; Gemini docs.
Le Chat “Memories” from Mistral
What it is: User-visible memory with on/off controls, log editing, exportability; scoped recall (“clickable receipts”) across chats, files and projects.
Why it matters: Memory done transparently reduces re-prompting, improves grounded answers, and addresses trust concerns.
Our take: The most interesting bit is agency - portable memories and clear provenance. Expect enterprise variants to compete with Microsoft’s Recall-style features.
Read more: The Rundown AI; Mistral
Sound Effects upgrades from ElevenLabs
What it is: SFX in Studio plus a new loop parameter (SDKs updated), enabling seamless ambience beds and precise durations from text prompts.
Why it matters: Makes AI audio production more production-ready... loopable atmospheres and more controllable SFX help video/podcast teams scale.
Our take: Combine with realtime VO to ship internal cuts without waiting on external libraries.
Read more: The Rundown AI; ElevenLabs changelog; SFX docs.
LongCat-Flash-Chat (560B MoE) from Meituan
What it is: A huge MoE model (≈27B active) reporting competitive reasoning and agent benchmarks; FP8 variant available.
Why it matters: Demonstrates cost-efficient scaling and strong agentic behaviors without “thinking” mode.
Our take: Watch its tool-use and long-context performance in real agent stacks.
Read more: The Rundown AI; Hugging Face.
Today’s Sources:
The Intenet