AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Trump Administration Orders Anthropic to Disable Claude Fable 5 and Mythos 5
A June 12 Commerce Department export-control directive forced Anthropic to immediately disable its top-tier Fable 5 and Mythos 5 models for all customers, citing national security concerns over offensive cyber capabilities.
US Data Center Bans and Moratoriums Exceed 300, With 75 More Under Consideration
State and local restrictions on data center development have surpassed 300 since 2023, with over 75 additional measures under consideration, as local opposition to the AI building boom spreads.
Key topics and reactions
Trump Administration Orders Anthropic to Disable Claude Fable 5 and Mythos 5
Anthropic released its Mythos-class models around June 9-10, 2026: Claude Fable 5 as the general-availability version with safety guardrails, and Claude Mythos 5 as a restricted-access tier. On June 12, the Trump administration's Commerce Department issued an export-control directive ordering access cut off for foreign nationals inside and outside the US, including Anthropic employees. Anthropic disabled both models for all customers immediately, while Claude Opus 4.8 and other models were unaffected.
The directive stemmed from the Mythos models' ability to automatically discover OS and browser vulnerabilities, capabilities shared with limited partners such as Amazon, Apple and JPMorgan Chase under a defensive security program called Project Glasswing. According to Wired, officials cited jailbreakable guardrails, model sharing with SK Telecom amid China-related suspicions, and concerns raised by Amazon CEO Andy Jassy.
Anthropic disputes the basis for the order, arguing the cited jailbreak method is narrow and not unique to its models, and that other public models have similar capabilities. As of June 26, no resolution was in sight. Anthropic reports Fable 5 scores 80.3% on SWE-Bench Pro versus 69.2% for Opus 4.8, priced at $10 per million input tokens and $50 per million output tokens.
OpenAI Releases GPT-5.6 Sol in Limited Preview Alongside Terra and Luna
OpenAI unveiled GPT-5.6 Sol as a limited preview, refreshing its lineup with a three-model structure: the high-end Sol, the balance-oriented Terra, and the speed- and cost-focused Luna. The company signaled further releases were on the horizon, while developers awaited concrete specifications such as pricing and context length.
According to The Guardian, OpenAI's decision to forgo a full public release reflects similar export-control requests from the US government, contrasting with rival Anthropic, which was forced to fully suspend its latest models.
Combined with Codex's mobile and cloud integration, developers said the staged rollout positions OpenAI's agent stack closer to handling planning through testing in a single workflow.
US Data Center Bans and Moratoriums Exceed 300, With 75 More Under Consideration
According to The Information, US state and local governments have imposed more than 300 bans or moratoriums on data center development since 2023, with over 75 additional measures in the pipeline. While generative AI demand has driven a surge in hyperscale construction by Amazon, Google, Meta and Microsoft, opposition has grown over power consumption of tens to hundreds of megawatts per facility, large cooling water use, noise and grid strain.
The National Conference of State Legislatures reports that as of June 2026, 14 states are considering data center bans or moratoriums. Tracker figures vary: datacenterbans.com counts 8 state-level measures in effect, while Interconnected Capital reports 97 enacted local moratoriums affecting 38 states.
Moratorium durations range from months to years, with examples including one year in Denver, through year-end in Oklahoma City, and three years in Michigan's Huron County. A federal bill, Senate bill S.4214, has proposed a moratorium on AI data centers, and Monterey Park, California passed a unanimous ban.
SpaceX, Groq and CoreWeave Expand AI Compute Leasing as Supply Diversifies
Google will pay SpaceX $920 million per month from October 2026 through June 2029 for compute equivalent to roughly 110,000 NVIDIA GPUs. Anthropic is reported to lease the full compute of the Colossus 1 data center near Memphis (over 220,000 NVIDIA GPUs) for $1.25 billion per month, while Reflection AI is leasing Colossus 2's NVIDIA GB300 chips for $150 million per month, up to $6.3 billion total.
Inference-focused Groq, following a roughly $20 billion licensing and technology-transfer arrangement with NVIDIA, has pivoted to its GroqCloud neocloud business and planned to raise $650 million from existing investors. Established players including CoreWeave, Lambda Labs, Crusoe, Nebius and Fluidstack continue expanding GPU clusters with cheaper, faster provisioning.
Neocloud on-demand H100 pricing can reach around $2.99 per GPU-hour, in some cases less than half hyperscaler rates, contributing to supply-side diversification alongside OpenAI's in-house chip development.
Category highlights
Foundation Model Roundup: Gemma 4, MiniMax M3, Nemotron 3 Ultra
Google's Gemma 4 31B reached 1,500 tokens per second on Cerebras as its first multimodal model. MiniMax M3 added NVFP4 support, and NVIDIA's Nemotron 3 Ultra performed well in 3D Design Arena. Mistral AI released lightweight models for edge devices, targeting AI use in resource-constrained environments.
Video Tools: Gemini Omni, OpenArt Vibe Directing, Vidu Multi-Entity
fal open-sourced 3DREAL (a render-to-real IC-LoRA for LTX-2.3). OpenArt's Vibe Directing enables consistent video up to five minutes from chat instructions, and HIX AI's Video Agent generates up to two-minute animations from a single prompt. Vidu AI added Multi-Entity Consistency, and Google announced Gemini Omni for editable cinematic video generation.
Audio: Suno K-pop Production and ElevenLabs ALS Voice Restoration
Suno showcased a case study producing a K-pop track in Seoul. ElevenLabs published a practical example of voice reconstruction for an ALS patient, highlighting accessibility applications of voice synthesis.
Platform: WebMCP Origin Trial, Figma Motion and Code Layers
The WebMCP web standard for AI agents entered origin trial, and Databricks rolled out Genie ZeroOps for operations monitoring. Figma added AI-driven motion graphics and code layers, while Microsoft 365 Copilot gained Excel personalization.
Events: I/O Connect Berlin, AI Engineer World's Fair, Runway Festival
Google held I/O Connect Berlin, including the Antigravity Agent SDK. The AI Engineer World's Fair, starting July 1, will feature Together and MiniMax. Runway's 2026 AI Festival winners were published, and Hugging Face Daily Papers for June 25 released 32 papers focused on video generation, agent systems, reasoning distillation and multimodal evaluation.
Key trends
Open-Weight Models Climb Arena Benchmarks
GLM-5.2, Nemotron 3 Ultra, MiniMax M3 and Gemma 4 are placing high on Arena benchmarks while balancing cost efficiency and performance, reinforcing a trend in which model selection is treated as a product strategy decision.