Live Feed
Nvidia in Talks to Acquire Reflection AI as Open-Weight Model Race Heats Up
The Financial Times reports Nvidia is in discussions to acquire or increase its stake in Reflection AI, maker of the 501B-parameter Beam model. A deal would mark Nvidia's first direct move into frontier model-making.
Manus Raises $500M+ from Boyu and IDG After China Blocked Meta's $2B Buyout
Butterfly Effect, the parent of AI agent startup Manus, closed more than $500 million from Boyu Capital, IDG, Tencent, and HSG six months after Beijing forced Meta to unwind its acquisition. The company has $100M ARR, just shipped Manus 2.0, and is weighing a Hong Kong IPO.
Inside Anthropic's $84.5B xAI Bet: The GOP Operator Who Broke a White House Deadlock
Tom Brown, Anthropic's Republican co-founder and chief compute officer, resolved a June White House model-safety standoff that CEO Dario Amodei couldn't break, securing $1.25B/month in xAI compute through May 2029 and anchoring the company's path to a $2 trillion IPO.
OpenRouter's Batch API Cuts Inference Costs 50% on 70+ Models — 7-Minute Median Turnaround in Beta
OpenRouter's Batch API sends an entire workload in one POST and returns results within 24 hours at half the per-token price. Two-week beta: 230k+ batches completed, median turnaround 7 minutes, 90th percentile under 1 hour.
Cloudflare Acquires Deno: Ryan Dahl's Team Merges workerd and celld for Self-Hosted Workers
The entire Deno team is joining Cloudflare to merge workerd with celld, Ryan Dahl's Rust-based distributed compute runtime. Goal: make the Workers programming model — the backbone of AI agent infrastructure — runnable anywhere.
Lenovo TianxiCode Tops SWE-bench-Live at 71% Running DeepSeek V4.1 Flash
Lenovo TianxiAI's TianxiCode code agent framework ranked first globally on SWE-bench-Live Lite with a 71% solve rate and Verified certification — using DeepSeek V4.1 Flash as its foundation model rather than a frontier reasoning model.
Gemini 4 Argon Leads Arena Text at 1,525 ELO and Posts the Lowest Frontier Hallucination Rate: 15%
Nine days after launch, Gemini 4 Argon tops Arena.ai's Text leaderboard at 1,525 ELO — ahead of Claude Opus 4.6 and Opus 5.5 — and posts the lowest hallucination rate of any frontier model: 15% on Artificial Analysis's AA-Omniscience, versus GPT-6 Astra max's 51%.
Asana's Browser Agent Runs 76x Cheaper on GPT-6.1 Sol After 144-Run Optimization Study
A 144-run head-to-head on four frontier models found GPT-6.1 Sol brought Asana's browser agent cost from $36 per run down to $0.47, with 89% of input tokens served from cache. Runtime dropped from 22 minutes to four.
Microsoft Decision-1 Is 35x Faster Than GPT-6 Sol on Structured Decision Tasks
Microsoft shipped Decision-1 through Azure Foundry and OpenRouter: a post-trained Qwen3.5-9B model for routing, classification, and prioritization that claims top accuracy across 36 blind benchmarks and P50 latency 35x below GPT-6 Sol.
TypeSafe AI Raises $870M at $7.5B — a Third of Fortune 500 Already Running Jev
Three weeks after launch, TypeSafe AI closed an $870M Series A led by Andreessen Horowitz and Sequoia, valuing the Jev model maker at $7.5B. The company says one in three Fortune 500 companies has already deployed the model.
GlobalFoundries Tops Out €1.1B Dresden Expansion and Launches FDX Fusion for Physical AI
GlobalFoundries reached a construction milestone on its €1.1B SPRINT expansion in Dresden and announced FDX Fusion, a next-generation FDSOI process platform designed for Physical AI with 7nm-like digital performance and integrated analog, RF, and embedded memory.
Oxide Computer Raises $445M Series D as AI Demand Pushes Enterprises Back to Owned Infrastructure
Oxide Computer has closed a $445M Series D led by Eclipse, with participation from Jane Street, citing profitability and enterprise demand for on-premises cloud hardware driven by AI compute workloads.
Iranian Operators Used ChatGPT to Plant ~100 AI-Generated Articles Across U.S. Publications
OpenAI banned ChatGPT accounts used by an Iranian influence operation that embedded roughly 100 AI-written articles critical of the U.S.-Iran war in at least 20 American news outlets, including outlets with millions of social media followers.
Microsoft Open-Sources MXC, a Cross-Platform SDK for Sandboxing AI Agent Workloads
Microsoft Execution Containers gives developers a unified API to run AI-generated code in isolated environments with declared filesystem, network, and UI policies. It supports eight containment backends across Windows, Linux, and macOS — and developers plug it into their own apps as an SDK dependency rather than running a separate daemon.
Three OpenAI Safety Researchers Fired, Dispute Misconduct Claims, Warn of Chilling Effect
Jasmine Wang, Tomek Korbak, and Mikita Balesni were dismissed last week after allegedly sharing confidential information with an external AI safety organization. In an open letter published Thursday, they denied the charges and warned that their firings are freezing the open culture OpenAI once relied on.