Live Feed
Claude Opus 4.8 Launches: Super-Agent Benchmark Leader, 84% on Mind2Web, Fast Mode 3x Cheaper
Anthropic released Claude Opus 4.8, the only model to complete every case end-to-end on the Super-Agent benchmark — beating GPT-5.5. Browser agent performance hits 84% on Online-Mind2Web. Fast Mode runs at 2.5x speed and is now three times cheaper than previous Opus fast pricing. Available today at Opus 4.7 pricing.
Anthropic Raises $65B in Series H at $965B, Overtakes OpenAI as Most Valuable AI Lab
Anthropic closed a $65 billion Series H led by Altimeter, Sequoia, Dragoneer, and Greenoaks, pushing post-money valuation to $965 billion — above OpenAI's $852 billion. Run-rate revenue crossed $47 billion this month. Micron, Samsung, and SK hynix join as strategic memory partners.
Cisco Tests 15 Frontier Models: Multi-Turn Attack Success Reaches 88% — Every Model Fails
Cisco's AI Threat Research team ran 30,090 single-turn and 6,986 multi-turn adversarial prompts against 15 closed frontier models from OpenAI, Anthropic, Google, Amazon, and xAI. Multi-turn attack success rates ranged from 7.9% to 88.3%. Not one model held.
Applied Digital's Mystery Hyperscaler Signed a $7.5B, 300MW Deal in Louisiana — Location Now Revealed
Applied Digital's Delta Forge 1 campus in Rapides Parish, Louisiana has been formally disclosed. The $3.6B site locked a $7.5B, 15-year take-or-pay lease with an unnamed investment-grade US hyperscaler in April — roughly $500M per year — covering all 300MW of initial capacity.
Nvidia Commits $150B a Year to Taiwan, Breaks Ground on Constellation Campus in Taipei
Jensen Huang announced at Computex 2026 that Nvidia's annual Taiwan spend will hit $150B, up 10x from five years ago. A new 4,000-person R&D campus called Constellation breaks ground this year in northern Taipei, targeting operations by 2030.
Cognition AI Raises $1B at $26B as Devin Revenue Climbs From $37M to $492M ARR in One Year
The autonomous coding agent company closed a $1B round at a $26B pre-money valuation, with ARR up 13x year-over-year. Goldman Sachs and Mercedes-Benz are in production. CEO Scott Wu pitches Devin as a model-agnostic agent layer routing across OpenAI and Anthropic rather than betting on one LLM.
The Grid Launches Quality-Tiered AI Inference Marketplace: Pick Standard, Prime, or Max, Not the Model
New inference router abstracts model selection into three quality tiers anchored to Artificial Analysis benchmarks. Developers send requests to text-standard, text-prime, or text-max; The Grid routes to the cheapest qualifying supplier and automatically ejects providers that drift below the quality floor.
SoftBank Launches Japan Sovereign AI GPU Cloud — Infrinia OS, AITRAS Edge, October Debut
SoftBank enters sovereign AI infrastructure with a GPU cloud bundled to its nationwide 5G network, targeting Japanese enterprises that won't route data through US-headquartered hyperscalers. Infrinia AI Cloud OS supports Kubernetes and LLM inference as a service. Launch: October 2026.
YouTube Stops Relying on Creators to Label AI Videos — Auto-Detection Rolls Out Platform-Wide
YouTube shifts from voluntary creator disclosure to automated detection for AI-generated content, labeling photorealistic AI videos without any creator action required. Existing labels get a visibility upgrade at the same time.
Codex Ran the Feedback Loop: OpenAI's Tax AI Processed 7,000 Returns at 97% Accuracy by Fixing Itself
OpenAI and Thrive Holdings built Tax AI for Crete's 30+ accounting firm network using a three-layer self-improvement loop: production traces, practitioner feedback, and a Codex agent that writes new evals and deploys fixes. At launch, only a quarter of returns hit 75% correct field completion. The system improved without a single engineering sprint.
DuckDuckGo iOS Installs Surge 70% After Google Makes AI Mode Default Search at I/O 2026
DuckDuckGo U.S. iOS app installs averaged 33% week-over-week growth for six consecutive days following Google I/O 2026, peaking at 69.9% on May 25. Visits to DuckDuckGo's AI-free search page climbed 27.7%. Google's AI Mode reports 1 billion monthly users — but 93% of those queries end without a click.
MiniCPM5-1B Leads Every Sub-2B Open-Weight Model by 7.4 Points on AA Intelligence Index
OpenBMB's MiniCPM5-1B scores 17.9 on Artificial Analysis Intelligence Index — 7.4 points ahead of the next-best sub-2B open-weight model and nearly 2 points above Alibaba's Qwen3.5 2B despite using fewer than half the parameters.
Meta/CMU Self-Play SWE-RL: Coding Agents Gain 10.4 Points by Training on Bugs They Made Themselves
A new paper from Meta, CMU and collaborators shows coding agents can manufacture their own training data — one agent injects bugs into real codebases, another repairs them. The result is +10.4 on SWE-bench Verified and +7.8 on SWE-bench Pro, with gains that transfer to natural-language issues the system never trained on.
China Locks Down AI Researchers: Alibaba and DeepSeek Engineers Need State Approval to Travel
Beijing has expanded travel restrictions to top AI talent at private firms, treating frontier model engineers as holders of sensitive national technology. Alibaba and DeepSeek staff are the first wave.
Mythos Found 1 Real Bug in curl. 360's Agent Tool Found 23 in Agent Systems. AI Security Has Split in Two.
A new divide in AI-powered security has emerged: model-based code scanners versus agent-native auditors. When pointed at the hardened curl codebase, Anthropic's Mythos surfaced 5 claimed vulnerabilities — only 1 survived expert review. A Chinese security team's agent-native tool found 23 critical flaws specifically because it reasons about tool chains, permissions, and hostile context as a coupled attack surface, not as static code.