Live Feed
Arena Adds Six Models Across Four Leaderboards in 48 Hours — Broadest Coverage Expansion of 2026
GPT-5.5-Instant enters the Text, Vision, and Document categories; ERNIE 5.1 enters Search; Gemma 4-31B and 26B-A4B enter Vision; Qwen 3.6 Max Preview and HunYuan Hy3-Preview enter Code. The two-day expansion gives the first apples-to-apples Arena comparisons for all six models.
OpenRouter Ships Unified Web Search and Fetch for Any Model — Four Engines, One API
OpenRouter's new server-side tools let any model call web search or page fetch without client-side plumbing. Exa integration costs $0.004 per search result; native provider search is included at provider pricing. The change removes the last major reason to wire search capability per-model.
RPCS3, Godot, SDL and llama.cpp Have All Banned AI-Generated PRs. GitHub Is Considering a Kill Switch.
Five major open-source projects formally banned undisclosed AI-generated pull requests in six months. GitHub opened a community discussion in February calling the volume a 'critical issue' and is evaluating tools to let maintainers block submissions outright.
Maryland Ratepayers Face $1.6B Decade Bill for Virginia's Data Centers — FERC Complaint Filed
Maryland's consumer advocate filed a FERC complaint on May 7, arguing PJM's $2B charge on the state is driven almost entirely by Northern Virginia data center demand. At stake: $1.6 billion in residential and commercial bills over ten years.
AI2's OLMo Hybrid Reaches Same Accuracy With 49% Fewer Training Tokens
Allen Institute for AI releases OLMo Hybrid, a 7B model combining transformer layers with Gated DeltaNet linear RNNs. It matches OLMo 3's MMLU accuracy using 49% fewer training tokens, with scaling law projections of 1.3-1.9x efficiency gains across model sizes from 1B to 70B parameters.
Anthropic Locks $1.8B Cloud Deal With Akamai — CDN Giant's Stock Jumps 27%
Anthropic has signed a $1.8 billion, seven-year infrastructure deal with Akamai, the largest contract in the CDN company's 28-year history. Akamai stock surged 27% on the news. The deal is Anthropic's third major compute commitment in weeks, as CEO Dario Amodei confirms 80x year-over-year growth in Claude usage.
NVIDIA Star Elastic: One Checkpoint, Three Deployable Model Sizes — 30B, 23B, and 12B
NVIDIA ships Star Elastic, a training method that embeds multiple nested submodels inside a single reasoning checkpoint. One training run on Nemotron Nano v3 produces 30B, 23B, and 12B variants that slice out at deployment time with no additional training. Benchmarks show 1.9x lower latency versus standard single-model control.
LLMorphism: New Paper Names the Cognitive Bias That Makes Humans Think Like LLMs
A preprint by Valerio Capraro introduces LLMorphism — the growing tendency for people to model their own cognition on LLM architecture. The paper argues the bias spreads through analogical transfer and linguistic metaphor, with downstream effects on education, healthcare, and legal responsibility.
DeepSeek Ships Image Recognition to All Users at 10x Fewer Tokens Than Rivals, V4.1 Due in June
DeepSeek has rolled out its image recognition mode to all users, processing an 800x800 image in roughly 90 tokens versus 870-1,100 for mainstream models. The company has briefed investors that V4.1 arrives in June.
Gemini API File Search Gets Multimodal RAG, Custom Metadata, and Page-Level Citations
Google has shipped three new capabilities to the Gemini API File Search tool: multimodal support for images and documents, custom metadata filters, and per-page citations. For enterprise RAG builders, page-level citations are the operative upgrade.
Isomorphic Labs in Talks for $2B+ to Scale AI Drug Discovery Engine Beyond AlphaFold
Google DeepMind's drug discovery spinoff is in advanced talks to raise more than $2 billion. Thrive Capital leads; Alphabet is participating. The money goes into IsoDDE, a drug design engine that doubles AlphaFold 3's score on the hardest binding-pocket tasks.
Arena's Seven Web Dev Categories Reveal Which Model Actually Wins Your Use Case
Aggregate Code Arena rankings hide significant model specialization. Claude Opus 4.7 Thinking leads broadly across all seven web dev categories; GPT-5.5 High dominates gaming and interactive simulations; Meta's Muse Spark tops brand marketing, informational websites, and consumer product apps.
IREN Locks $3.1B ARR, Moves Into Spain, and Acquires Mirantis — the Bitcoin Miner's AI Pivot Is Complete
IREN's Q1 2026 update reveals $3.1B in contracted annual recurring revenue against a $3.7B year-end target. Two acquisitions — Nostrum (490MW in Spain, 1GW+ pipeline) and Mirantis (software orchestration) — extend the company's reach beyond its US GPU farms. 1,210MW of 2027 capacity is now under construction.
ERNIE 5.1 Launches at 6% of Industry Pretraining Cost — Arena Search #4 Globally, Beats DeepSeek-V4-Pro on Agents
Baidu's GA release compresses ERNIE 5.0 to one-third the total parameters, slashes pretraining compute to 6% of comparable models, and lands at 1,223 Elo on the Arena Search leaderboard — the only Chinese model in the global top four.
NVIDIA Has Bet $40B on AI Equity Deals in 2026 — Most of It Flows Back to Its Own Customers
The chipmaker has committed more than $40 billion to equity investments in AI companies in the first five months of 2026 — $30 billion of it in OpenAI — plus seven multi-billion deals in public companies and roughly two dozen private rounds. Wall Street calls it circular. Jensen Huang calls it a moat.