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
Key topics and reactions
Anthropic Halts Claude Fable 5 and Mythos 5 After White House Export Directive
Anthropic released Claude Fable 5 and the more capable, guardrail-free Claude Mythos 5 around June 9, 2026. Fable 5 was offered via the Claude API as claude-fable-5 at $10 per million input tokens and $50 per million output tokens, and briefly outperformed GPT-5.5 on several coding and reasoning benchmarks at launch.
On June 12, the Trump administration issued an export-control directive over cybersecurity concerns, including potential discovery of unknown vulnerabilities and jailbreak risks. The directive restricts access for foreign nationals, including Anthropic's own staff, and the company immediately disabled both models for all users while leaving other Claude models unaffected.
OpenAI has reportedly been asked by the government to stagger or limit the release of GPT-5.6 over similar security concerns. The administration, which previously favored deregulation, is increasingly exercising pre-release review over frontier model launches.
Claude Opus 4.8 and Haiku 4.5 Reach General Availability on Microsoft Foundry
Anthropic announced on June 29, 2026 that Claude Opus 4.8 and Claude Haiku 4.5 are generally available on Microsoft Foundry, hosted on Azure. Both run via Anthropic's Messages API and support prompt caching and extended/adaptive thinking. The deployment runs on NVIDIA GB300 NVL72 with Quantum-X800 InfiniBand.
The integration supports Microsoft Entra ID or API key authentication, Claude Consumption Units billing through Azure Marketplace, and consumption from existing Azure Enterprise Agreement commitments, letting enterprises run Claude in production without new procurement. Deployment is offered in Global Standard regions such as East US 2 and Sweden Central, with Opus 4.8 also supporting a US Data Zone Standard.
Opus 4.8 offers a 1M-token context window and 128K maximum output, while Haiku 4.5 provides 200K context and 64K output. Both support effort-level control, tool use, high-resolution image input and streaming, though the Azure-hosted Version 2 carries some feature limits compared with Anthropic's direct API.
NVIDIA Releases Open-Weight Omni Model Cosmos 3 With Cross-Modal Reasoning
NVIDIA released Cosmos 3, which it describes as the first open-weight omni model with native visual reasoning across text, image, video and audio. The 64B Cosmos 3 Super is available on Design Arena.
The release adds to a wave of open-weight, multi-modal models gaining ground on both quality and modality coverage. NVIDIA also previewed its SIGGRAPH 2026 Research Keynote, focused on neural rendering and world models.
The model positions open releases as competitive across modalities at a time when proprietary frontier models face new regulatory scrutiny.
Cursor and OpenClaw Launch Mobile Apps for Cloud Coding Agents
Cursor's iPhone app lets users draft, review and merge pull requests entirely from a phone, with a workflow of dispatching tasks remotely and checking results via notifications. Limits noted include no EU availability and the difficulty of writing long prompts and reviewing complex multi-file edits on a small screen.
OpenClaw uses a Telegram-style channel to maintain continuous interaction with an agent, automating email, calendar, Linear and support inbox management. Some users report the agent handling spam, scheduling and to-dos after a 10-14 hour setup, while others cite missing tools requiring custom CLIs, poor token efficiency, and unstable behavior on models other than Opus.
The launches reflect a broader move toward always-on, location-independent agent operation.
Category highlights
Video Generation: LTX-2.3, Kling, Netflix Research and Wan Streamer
LTX-2.3 took the top open-weight spot on Video Arena with an Elo of 1138, up 115 from the prior model. Kling made its Motion Video Generation generally available, while Netflix Research presented the Vera layered video diffusion model and the physics-aware inpainting system VOID. Google Vids added Veo-powered parallel generation and improved consistency, and Alibaba's Wan Streamer demonstrated 25 FPS low-latency real-time interactive video. PixVerse promoted Seedance 2.0 native 4K cinematic generation.
LangChain Integrates NVIDIA Nemotron Models Across Agent Workflows
LangChain integrated NVIDIA's open Nemotron models across agent workflows from reasoning to orchestration, exposing them as a production-ready open stack. Developers can call Nemotron on LangChain, LangGraph and Deep Agents via the ChatNVIDIA class in the langchain-nvidia-ai-endpoints package. Models are served as OpenAI-compatible APIs on NVIDIA NIM microservices, deployable through the hosted NVIDIA API Catalog or self-hosted with an NVIDIA AI Enterprise license. LangChain added Day 0 support for the 550B-parameter, 55B-active MoE Nemotron 3 Ultra around June 4, 2026.
Design Arena Adds Video-to-Website Generation
Design Arena introduced a Video-to-Website feature on June 29, 2026 that takes video and text as input to generate dynamic, high-fidelity sites, reflecting motion, timing and visuals into animations and interactions. Output supports code download, publishing and editing, with a comparison leaderboard coming soon. The platform, run by The Intelligence Company and a YC S25 alum, gives users free access to top models including Claude, GPT, Gemini and Grok, with Elo-style voting, and reports over 4.8 million users.
Voice and Music: Grok APIs and ElevenLabs Upgrades
Grok voice APIs (TTS and STT) entered beta on Vercel AI Gateway. ElevenLabs upgraded its multilingual voice cloning with stronger emotion control and is deploying outbound recruiting voice agents via ElevenAgents.
Databricks Announces Multi-Agent Meta-Harness Omnigent
Databricks announced Omnigent, an open meta-harness for integrating multiple agents. Alongside agent evaluation benchmarks like Claw-Eval, such integration layers are making agent interoperability and long-horizon task performance a new competitive axis. Arena reached a $100M annual run rate eight months after launch.
Key trends
Foundation Models: StepFun Step 3.7 Flash Ranks Second on Claw-Eval
StepFun's Step 3.7 Flash placed #2 on Claw-Eval General, behind Claude Opus 4.6, performing well on long-horizon tasks. The result adds to a field where open-weight omni models like NVIDIA Cosmos 3 are pushing on both modality coverage and quality.