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 Confidentially Files Draft S-1 with SEC, Setting Up Possible IPO
Anthropic, PBC announced on June 1, 2026 that it filed a draft S-1 registration statement with the U.S. Securities and Exchange Commission on a confidential basis, aimed at an eventual initial public offering. The number of shares and price range remain undetermined, and the company stressed the disclosure is made under Rule 135 of the Securities Act and is not an offer to sell securities.
The move follows a sharp rise in Anthropic's valuation. On May 28, 2026 the company raised $6.5 billion in a Series H round at a post-money valuation of roughly $965 billion, which Reuters reported makes it the world's most valuable AI company, surpassing OpenAI's most recent valuation. That valuation is up about 2.5x from its Series G just months earlier.
On the same day as the filing, Anthropic also announced Claude Opus 4.8 for coding and agentic tasks.
OpenAI Frontier Models and Codex Reach General Availability on AWS Bedrock
OpenAI and AWS announced on June 1, 2026 that OpenAI's frontier models—including GPT-5.5 and GPT-5.4—plus Codex and Amazon Bedrock Managed Agents powered by OpenAI are now generally available on Amazon Bedrock. The services launched as a limited preview on April 28, 2026 and have now moved to GA.
Codex can be used through Bedrock with the Codex CLI, desktop app, and VS Code extension. Usage runs through the existing Bedrock API, with AWS security, governance, and cost management applied automatically and consumption counting toward existing AWS cloud commitments, so no separate security model is required.
The launch marks the first time OpenAI models are natively available on a major cloud outside Microsoft Azure, following the April 2026 restructuring of the Microsoft–OpenAI exclusivity arrangement.
xAI Integrates Cursor's Composer 2.5 Model into Grok Build
xAI announced on June 1, 2026 that Composer 2.5 is now selectable within Grok Build, describing it as a fast and highly capable model that excels at long-horizon tasks and following complex instructions. Composer 2.5 is Cursor's in-house coding model, released on May 18, 2026 and built on a checkpoint of Moonshot's open-source Kimi K2.5 with large-scale reinforcement learning improvements.
Grok Build is xAI's Rust-based coding agent CLI, offering up to eight parallel sub-agents and a Plan mode, and is available to SuperGrok and X Premium Plus subscribers. Early users on X reported building functional apps quickly, with one claiming a fully designed global recipe app in ten minutes.
Some users flagged heavy consumption, with reports of burning through about 20 days' worth of usage at once and uncertainty over whether Composer 2.5 usage is shared or separately metered, suggesting heavier tiers such as SuperGrok Heavy may be needed for sustained work.
MiniMax Launches Open-Weight M3 with 1M Context and Frontier-Class Coding
MiniMax M3 launched as an open-weight model combining frontier-class coding, long context, and multimodal abilities. It appeared on OpenRouter the same day and was made available on Arena—for Text, Vision, Document, and Code (Frontend) categories—with scores to be published soon.
Developers cited concrete results, including 59% on SWE-Bench Pro, 145 CUDA-operator iterations over 24 hours, and practical agentic tasks at 1M context. Several called it a game-changer for Chinese open-weight efforts to match closed frontier models.
Reported limitations include occasional Chinese-language output drift during long runs and a wait of roughly ten days before the official weights are released.
Category highlights
Claude Opus 4.7 Dominates Design Arena Slides
Anthropic's Claude Opus 4.7 and Claude Opus 4.7 (Thinking) took the top two spots in Design Arena's Slides category, leading other models by more than 80 Elo points, according to the benchmark's June 1 post. Z.ai's GLM 5.1, Google DeepMind's Gemini 3.5 Flash, and Moonshot's Kimi K2.6 followed, with Anthropic effectively setting the standard for agentic slide generation.
ChatGPT Adds Full-Screen Editing and Library Save
OpenAI's ChatGPT account announced a new full-screen editing experience for long-form writing, letting users compose and edit in a wider space and save documents to the Library to resume later. The feature, demonstrated in a clip, continues ChatGPT's shift from Canvas toward Writing/Code Blocks and the persistent Library, which offers per-tier storage limits and search and Memory integration.
Voice-First 'Agentic OS' Wins OpenAI Hackathon
OpenAI announced that solo developer @isausmanov's "Agentic OS for a Phone" won the People's Choice vote at the May 27 Voice Hack Night in San Francisco, earning $50,000 in OpenAI API credits. The voice-first mobile OS prototype, built in about six hours, lets users speak to an agent that acts across phone apps and features, echoing reported OpenAI work on an agent-centric phone.
New Agent-Oriented Search and Memory Tools
Perplexity announced "Search as Code," a search architecture aimed at AI agents, while Tencent released Hy-Memory, an agent memory plugin from its Hunyuan team. The releases reflect a broader push toward infrastructure optimized for long-horizon, tool-using agents alongside this week's wave of agentic models.
NVIDIA Robotics Unveils Open Humanoid Reference Design
NVIDIA Robotics announced its first open humanoid robot reference design, developed jointly with partners. The move extends NVIDIA's push into physical AI and robotics, complementing Runway's launch of a Cosmos Coalition with NVIDIA and others around world models.
Key trends
Open-Source Talking-Avatar Tools Cut Production Costs
Open-source projects including SadTalker, MuseTalk, LivePortrait, and Hallo2 can generate high-fidelity talking avatars from a single image plus audio, positioning them as a foundation for real-time AI agents. Chinese open-source tools similarly produce lip-synced video from one photo and audio, sharply lowering production costs.
Chrome Opens HTML-in-Canvas API Origin Trial
Google began an origin trial in May 2026 for an experimental HTML-in-Canvas API that renders real HTML/DOM inside a <canvas> element, available in Chrome 148–150. It aims to bridge Canvas's low-level graphics with DOM interactivity and accessibility, with primitives like the layoutsubtree attribute and drawElementImage(), and experimental support from libraries such as Three.js and PlayCanvas.