Live Feed
Claude Fable 5 Voids Enterprise Zero-Data-Retention Agreements: Safety Surveillance Is the Price of Frontier Access
Anthropic is mandating 30-day traffic retention for all Fable 5 and Mythos 5 users, overriding previously negotiated zero-retention agreements. AWS Bedrock requires explicit opt-in; GitHub Copilot defaults off. The policy extends to all future models at comparable or higher capability levels.
Anthropic Ships Claude Fable 5: First Mythos-Class Model Goes Public at 80.3% SWE-Bench Pro
Fable 5 scores 80.3% on SWE-Bench Pro — 11 points above Opus 4.8 and 22 above GPT-5.5. It is the same underlying model as Mythos 5, wrapped in safety classifiers that route fewer than 5% of sessions to Opus 4.8. Priced at $10/$50 per million tokens.
Google Ships Gemini 3.5 Live Translate: Continuous Speech-to-Speech in 70+ Languages, SynthID-Watermarked
Unlike turn-by-turn systems, 3.5 Live Translate generates translated speech as audio streams in, staying a few seconds behind the speaker. Rolling out today to Google Translate on iOS and Android, Google Meet enterprise preview, and Gemini Live API for developers.
Broadcom Launches $35B AI XPV Platform With Apollo and Blackstone — 20GW of Custom Silicon by 2028
Broadcom, Apollo, and Blackstone created a private credit vehicle to finance 20+ gigawatts of XPU-based compute by 2028. The first tranche is $35B, anchored by Anthropic's 1GW+ expansion at Fluidstack sites. OpenAI is named as a future platform customer.
MiniMax M3 Tops Open-Weight Intelligence Index at 55 — Weights Still Missing on Day 8
Artificial Analysis independently confirmed MiniMax M3 at Intelligence Index 55 on June 8, making it the leading open-weight model ahead of Kimi K2.6 at 54. The company promised weights at launch on June 1. They have not shipped.
UK Puts £1.1bn Into AI Chips at London Tech Week — Google, Anthropic, Microsoft, OpenAI Sign Government Pact
Starmer's June 8 London Tech Week speech committed £1.1bn to a national AI hardware plan including a £750m AI supercomputer and £400m in chip purchases, while all four major frontier labs signed a joint statement backing UK government AI policy.
Anthropic Names Its Mythos Public Version: Claude Fable Lands as Early as June 10
Anthropic's Mythos-class model goes public under the name Claude Fable as early as tomorrow, with new research showing the model turns N-day vulnerabilities into working exploits in under an hour. The N-day threat window has collapsed to N-hour.
Databricks Eyes $175B Valuation on $5.4B ARR — The Data Layer Is Now Worth More Than Most Foundation Model Labs
Databricks is in discussions for a new round at $165-175B, three months after closing at $134B. Revenue hit $5.4B ARR in February at 65% growth, with AI products generating $1.4B alone. The data infrastructure layer is pricing itself above every AI lab except Anthropic and OpenAI.
Congress Drops First Comprehensive US AI Law: 3-Year State Preemption, $500M Revenue Threshold, $300M CAISI
A bipartisan House discussion draft released June 4 proposes the most ambitious federal AI framework yet: overriding state development rules for three years, mandating risk reporting from companies above $500M revenue, and giving CAISI $300M to run independent audits.
Moonshot AI Seeks $30B in Its Third Raise in Six Months — Kimi ARR Doubled to $200M in April
The Beijing-based developer of Kimi has opened talks for up to $2B at a $30B valuation, its third financing round since December. That is a 7x jump from its $4.3B December valuation on $200M ARR, and it slots Moonshot alongside DeepSeek and Zhipu in China's top-funded AI tier.
Krea Launches First Foundation Image Model at No. 6 on Artificial Analysis, Trading Photorealism for Creative Direction
KREA-2 debuted at No. 6 on the Artificial Analysis Text-to-Image Leaderboard with Arena ELO of 1,121. Krea's first model trained from scratch deliberately prioritises style control and moodboard consistency over prompt accuracy or photorealism.
HRM-Text Trains a Competitive 1B Model for $1,500, Matching 7B Transformers on 100x Fewer Tokens
Sapient Intelligence's hierarchical recurrent architecture achieves 84.5% GSM8K and 60.7% MMLU at 1B parameters using just 40B training tokens. Independent verification confirms results hold under contamination-free conditions — and a HuggingFace Transformers PR is open.
ChatGPT's Web Share Fell 22 Points in 15 Months. Gemini Now Has 27%. Claude's Traffic Is Up 306%.
Momentic's Similarweb analysis of seven AI assistants, published June 8, shows the generative AI market shifting from near-monopoly to three-way race. The numbers capture web visits only, but the trend is structural: ChatGPT is declining faster on this metric than it rose.
Microsoft Research Lens: 3.8B Image Model Beats 80B Rivals at One-Fifth the Training Compute
Microsoft Research's open-source Lens model uses 800M GPT-4.1-captioned images and a 3.8B architecture to beat Hunyuan-Image-3.0 (80B) across prompt fidelity and text rendering benchmarks. Lens-Turbo generates a 1024x1024 image in 0.84 seconds.
Claude Opus 4.8 Hits 1512 on Chatbot Arena — First Model to Break the 1510 Barrier
Two days after joining the Arena text leaderboard on June 6, Claude Opus 4.8 settled at 1512 ELO — seven points clear of the nearest rivals and the first model to breach 1510. On coding, it leads at 1582, ahead of Opus 4.7 at 1567.