Anthropic files its IPO paperwork, NVIDIA launches Cosmos 3, MiniMax M3 and Qwen3.7-Plus

June 1, 2026 will remain one of the densest days in recent AI history: Anthropic takes a step toward the stock market with a confidential S-1 filing with the SEC, while a “Physical AI” wave rolls out simultaneously — NVIDIA Cosmos 3, Runway+NVIDIA Cosmos Coalition, Luma OPAL Lab, MiniMax M3, Qwen-VLA, and Qwen3.7-Plus. At the same time, OpenAI is deploying Rosalind Biodefense for pandemic preparedness and publishing a governance framework aligned with California’s TFAIA and the EU AI Act.

Anthropic files a confidential S-1 with the SEC

June 1 — Anthropic has confidentially filed a draft registration statement on Form S-1 with the U.S. Securities and Exchange Commission (SEC) in preparation for an initial public offering (IPO) of its common stock. This confidential filing is a mandatory preliminary step that allows the SEC to review the document before any public offering.

Item	Value
Filing type	Confidential S-1 form
Regulator	SEC (United States)
Share type	Common stock
Price / number of shares	Not yet set
Legal status	Preliminary filing — not an offer to sell
Date	June 1, 2026

The announcement states that Anthropic retains the option to go public once the SEC review is complete, but that the final transaction will depend on market conditions and other factors. This S-1 filing comes in direct continuation of the close of Series H (USD 65 billion, post-money valuation of USD 965 billion) and an annualized revenue figure that reportedly surpassed USD 47 billion in mid-May 2026.

Anthropic remains a public benefit corporation (PBC — Public Benefit Corporation), a status it intends to keep after a possible listing — a rare legal distinction that imposes public-interest obligations in addition to the pursuit of profit.

🔗 Official announcement — Anthropic

NVIDIA Cosmos 3 — first open-source omni-model for Physical AI

June 1 — NVIDIA launches Cosmos 3, the first fully open omni-model for physical AI. It unifies language, image, video, audio, and action generation in a single model — a world first.

Variant	Parameters	Availability
Cosmos 3 Super	32B	Open-source at launch
Cosmos 3 Nano	8B	Open-source at launch

The model natively integrates visual reasoning (native vision reasoning), world generation (world generation), and action generation (action generation). According to Artificial Analysis, Cosmos 3 immediately ranks #1 among open-weight models for text-to-image and image-to-video generation on day one. The weights are available on Hugging Face and NVIDIA NGC.

“Introducing Cosmos 3: Our latest frontier model for Physical AI. Cosmos 3 is the world’s first fully open omnimodel with native vision reasoning, world and action generation.” — @NVIDIAAI on X

Cosmos Coalition — Runway + NVIDIA, a global initiative for world models

June 1 — Runway joins NVIDIA as a founding member of the Cosmos Coalition, a global initiative bringing together leading AI labs to build and make open-source frontier world models for physical AI.

The goal is to create a shared open ecosystem allowing any research or industry group to use, modify, and assemble these models for productive work systems. The coalition also includes partnerships with chip and hardware manufacturers.

“Introducing the Cosmos Coalition. A new global initiative with NVIDIA and leading AI labs to build and open-source frontier world models for physical AI.” — @runwayml on X

This coordinated launch with Cosmos 3 marks a strategic turning point: NVIDIA is positioning open-source physical AI as an issue of global technological sovereignty, in contrast to the concentration of resources among a few closed players.

MiniMax M3 — first open-weight model to combine frontier coding, multimodality, and 1M context

June 1 — MiniMax launches M3, the first open-weight model combining three frontier capabilities simultaneously. With 1.7 million views in just a few hours, it is one of the most-followed open-source launches of the year.

Capability	Detail
Coding & Agentic frontier	59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1
1M token context	MiniMax Sparse Attention
Native multimodality	From the very first training stage

Full benchmarks:

Benchmark	Score
SWE-Bench Pro	59.0%
Terminal Bench 2.1	66.0%
SWE-fficiency	34.8%
KernelBench Hard	28.8%
MCP Atlas	74.2%
BU Bench (browser use)	+26% vs previous

M3 is natively designed for computer-use agents and long-duration agentic workflows. The weights and technical report will be released in about 10 days from launch. A dedicated IDE, MiniMax Code, is launched simultaneously, accessible at code.minimax.io. First-day launch partners include Qubrid AI (GPU cloud infrastructure) and Simular (autonomous agents).

🔗 MiniMax M3 announcement

Luma OPAL Lab — open Physical AI lab for general-purpose robotics

June 1 — Luma announces the creation of the OPAL Lab (Open Science Physical AI Lab), an open research laboratory dedicated to generalization in robotics.

Founder Amit (@gravicle) identifies the central problem in current physical AI: unlike large language models capable of generalizing to novel tasks, robots can only reproduce specific tasks in specific environments. Robotics suffers from a data shortage that creates a generalization crisis.

“We are launching Open Physical AI Lab at Luma to solve generalization in robotics. We believe general physical AI is the most consequential problem to solve to improve life on Earth.” — @gravicle on X

Luma is building on 4 years of fundamental research in multimodal AI (3D, image, video, unified generation) and on internet-scale multimodal data processing infrastructure. The OPAL Lab principles: open science, methods and substrates available to everyone, collaboration with academia, industrial partnerships, and safety evaluations. The message is explicitly political: concentrating the world’s physical intelligence infrastructure in the hands of a small number of companies represents a profound threat to humanity.

🔗 OPAL Lab announcement

Qwen3.7-Plus — Alibaba’s multimodal agent with vision, coding, and long context

June 1 — Alibaba Tongyi Lab releases Qwen3.7-Plus, its most advanced multimodal agent model. Built on the Qwen3.7 text backbone, it merges vision and language into a single, versatile agent foundation.

Key benchmarks:

Benchmark	Qwen3.7-Plus	GPT-5.4	Claude Opus 4.6 Max	Gemini 3.1 Pro
Terminal Bench 2.0	70.3	—	65.4	—
SWE-bench Verified	77.7	—	80.8	—
GPQA Diamond	90.3	—	91.3	—
ScreenSpot Pro	79.0	67.4	49.5	68.1
AndroidWorld	81.0	—	62.0	70.7
BabyVision	70.4	53.1	12.6	55.9

Four main capabilities: hybrid GUI+CLI agent (perceives screens, navigates applications), coding and multimodal productivity assistant (from frontend mockups to complex engineering), visual agent (perception, reasoning, localization, and web-search-augmented Q&A), and multi-environment generalization via Claude Code, OpenClaw, Qwen Code, or any other agent environment.

Availability: Alibaba Cloud Model Studio, OpenAI-compatible API, and Anthropic protocol (usable directly with Claude Code via ANTHROPIC_BASE_URL).

🔗 Qwen3.7-Plus blog

Rosalind Biodefense — OpenAI opens GPT-Rosalind to defensive biology

May 29 — OpenAI announces two major initiatives around defensive biology, built on GPT-Rosalind, its reasoning model dedicated to the life sciences.

Organization	Field	Type
Fourth Eon	Biosafety screening, DNA synthesis	Rosalind Biodefense developer
SecureDNA	Biological security	Rosalind Biodefense developer
Lawrence Livermore National Laboratory	Medical countermeasures	Government partner
Johns Hopkins APL	Protein engineering, therapeutics	Government partner
CEPI	Pandemic vaccines, Ebola response	Government partner

Rosalind Biodefense is a development program open to academic institutions, nonprofit organizations, mission-driven defensive companies, and government teams. OpenAI funds access to GPT-Rosalind and supports projects on epidemiological modeling, early detection, screening, pandemic preparedness, and medical countermeasures.

“Advances in biology can strengthen our ability to prevent, detect, and respond to biological threats. Our goal is to help build a more robust ecosystem — giving trusted defenders frontier AI to develop and operate new defenses for public health and biodefense.” — @OpenAI on X

Expanded access to GPT-Rosalind for U.S. and allied government partners is announced simultaneously, with beneficiaries such as CEPI as part of its “100 Days” vaccine acceleration mission — including the response to the ongoing Ebola outbreak.

🔗 Rosalind Biodefense — OpenAI

Claude Code v2.1.153 to v2.1.159 — five updates after May 28

May 29 – June 1 — Anthropic releases five versions of Claude Code in the days following the launch of Opus 4.8.

Version	Key points
2.1.153	`skipLfs` option for marketplace, improved `claude doctor`, MCP/macOS/Windows agent fixes
2.1.156	Critical fix — thinking blocks in Opus 4.8 were causing API errors
2.1.157	Auto-loaded plugins from `.claude/skills/`, `plugin init` command, `EnterWorktree` mid-session
2.1.158	Auto Mode on Bedrock, Vertex, and Foundry for Opus 4.7 and Opus 4.8 (`CLAUDE_CODE_ENABLE_AUTO_MODE=1`)
2.1.159	Internal infrastructure, no visible changes

Version 2.1.157 is the most notable: plugins placed in .claude/skills/ are now loaded automatically without going through the marketplace, and EnterWorktree can switch between Claude-managed worktrees during a session. Version 2.1.158 brings Auto Mode to the major cloud platforms, strengthening Claude Code’s enterprise deployment.

🔗 Claude Code releases

Gemini 3.5 Flash and Gemini Omni generally available

May 29 — Google marks the general availability (general availability) of Gemini 3.5 Flash through 9 video demos covering Gemini Omni and Gemini 3.5 Flash in action.

Gemini 3.5 Flash rollout: default model in the Gemini app and in Search AI Mode (global), available via Google Antigravity, the Gemini API in Google AI Studio, Android Studio, Gemini Enterprise Agent Platform, and Gemini Enterprise.

Gemini Spark: the 24/7 personal AI agent powered by Gemini 3.5, now available to all Google AI Ultra subscribers in the United States, integrated into Gmail, Docs, Slides, and Workspace.

Gemini Omni Flash: rolled out to Google AI Plus, Pro, and Ultra subscribers in the Gemini app and Google Flow; available free of charge for YouTube Shorts and YouTube Create App users; developer API rollout planned in the following weeks.

🔗 Gemini Omni and 3.5 Flash demos

Qwen-VLA — vision-language-action robotic agent for 11 types of robots

May 29 — Tongyi Lab releases Qwen-VLA, a unified Vision-Language-Action (VLA) model for general embodied intelligence (general embodied intelligence).

Architecture: Qwen3.5-4B (vision-language backbone) + 1.15B DiT decoder to produce actions directly from visual perception and natural language. The model unifies object manipulation, spatial navigation, and trajectory prediction in a single framework.

Thanks to embodiment-aware prompts (embodiment-aware prompts), the same model runs on 11 robotic embodiment types — single-arm, dual-arm, and humanoid platforms — without policy heads or task-specific architectures. The tweet generated 3.1 million views in two days, a sign of major interest from the robotics community.

🔗 Qwen-VLA on X

Copilot — the metrics API adds AI adoption cohorts

May 29 — GitHub is enhancing the Copilot usage metrics API with an AI adoption cohort system. Each active user is classified into one of four phases below, calculated over a rolling 28-day window:

Phase	Name	Criterion
0	No cohort	Not enough engagement
1	Code first	Code completion and/or IDE agent mode
2	Agent first	One GitHub agent surface (cloud agent, code review, CLI)
3	Multi-agent	Two or more agent surfaces, or the GitHub Copilot app

Metrics exposed by phase: active users, interactions, code generation and acceptance, lines added/deleted, pull requests created/merged/reviewed, median merge time. Accessible to enterprise admins and organization owners. This system makes it possible to measure the real maturity of Copilot adoption and track developers’ progress toward advanced agentic use cases.

🔗 GitHub Changelog — Copilot usage metrics API

ElevenLabs signs an MOU with the Government of Greece

May 28 — ElevenLabs signs a memorandum of understanding with the Greek government, in the presence of the Prime Minister (@PrimeministerGR) and the Minister of Digital Transformation (@papastergiougr), around three priorities: improving access to public services via voice AI, promoting tourism, and preserving Greek linguistic heritage. This government partnership illustrates ElevenLabs’ strategy to establish itself in the European public sector by focusing on highly visible institutional use cases.

🔗 ElevenLabs announcement

Frontier Governance Framework — OpenAI aligns its practices with the TFAIA and the EU AI Act

May 28 — OpenAI publishes its Frontier Governance Framework, a public governance document that translates the internal Preparedness Framework into concrete regulatory obligations. The main target: California’s Transparency in Frontier AI Act (TFAIA) and the EU AI Act Code of Practice for general-purpose AI.

The framework covers assessment and mitigation of risks in offensive cyber, CBRN (chemical, biological, radiological, nuclear), harmful manipulation, and loss of control. It also includes model reporting obligations, incident management, use of external experts, and an update mechanism. OpenAI specifies that the Preparedness Framework remains its internal foundation — some practices go beyond current legal requirements.

🔗 Frontier Governance Framework

Trusted third-party evaluations guide — method and five biases to control

May 29 — OpenAI publishes a methodological guide for rigorous third-party evaluations of frontier AI models. The document starts from a simple observation: evaluations originally designed for chatbots are no longer suited to agentic systems capable of using tools and acting over long trajectories.

The guide distinguishes three types of claims that evaluations must support: capability elicitation (what the model can do under the best conditions), robustness of safeguards (resistance to attacks), and comparison across systems. A central concept is the “harness” (the execution environment around the model), whose choices can materially change results — OpenAI cites the impact of context compaction on GPT-5.5 performance in multi-step cyber evaluations.

Five validity biases are identified: reward hacking (exploiting shortcuts), refusals masking real capabilities, contamination (benchmark memorization), broken problems (impossible or poorly scored tasks), and sandbagging (deliberate underperformance in detected evaluations). OpenAI commits to providing maximal elicitation instructions to third-party evaluators and to sharing reasoning traces.

🔗 Trusted third-party evaluations guide — OpenAI

Grok Imagine Extend — continuous video extension

May 29 — Grok Imagine launches Extend, a feature that lets users create continuous videos with smooth transitions from an existing video or images. The feature is available directly in the Grok Imagine interface. The tweet from the @imagine account was reposted by @grok and reached 701,000 views. This announcement strengthens Grok’s creative suite alongside Grok Build and the grok-build-0.1 API.

🔗 Grok Imagine Extend

Codex Computer Use on Windows and mobile control

May 29 — OpenAI announces that Codex’s “computer use” feature is now available on Windows. The agent can see the screen, click, and type in Windows applications, opening the door to native testing, simulator workflows, and bug fixes accessible only through a graphical interface. The ChatGPT mobile app simultaneously adds Windows support for Codex: start, monitor, and control tasks from the phone while the work continues on PC. OpenAI describes this feature as an “early experience”; it is not available in the European Economic Area, the United Kingdom, or Switzerland.

🔗 Codex Computer Use on Windows

Briefs

Salesforce + Claude Code — Boris Cherny (@bcherny) shares a Salesforce testimonial on the agentic use of Claude Code: a migration initially estimated at 231 days delivered in 13, 21 endpoints at 100% test coverage, and a 5% drop in incidents despite an increase in PR volume. 🔗 source
Google I/O 2026 — AI behind the scenes — Google explains how its teams used Gemini Omni, Nano Banana, Lyria 3, and Antigravity to create the creative experiences for Google I/O 2026: animated short film, visual identity, interactive musical pre-show, generative games. 🔗 source
Gemini CLI v0.44.0 — Unified Auto Mode (merging specialized modes), native support for Sublime Text and Emacs Client, new programmable TUI testing tools. 🔗 source
GitHub Copilot — models under evaluation in auto mode — Individual users (Free, Pro, Pro+) now have access to pre-GA models through the automatic selection auto. A disable option is available in settings. 🔗 source
ElevenLabs Summit Warsaw — Preview of the most expressive voice model ever developed by ElevenLabs, unveiled by co-founder Mati. This is a preview only, with no public access announced. 🔗 source
Cohere Command A+ — translation benchmarks — Command A+ outperforms Mistral Medium 3.5, DeepSeek, gpt-oss, and Claude Opus 4.6 on WMT24++, as well as Google Translate. Documented gains: +2.4 pts in French, +1.9 pts in Spanish, +0.9 pts in German. 🔗 source

What this means

Anthropic’s financial momentum and sector consolidation. Anthropic’s confidential S-1 filing is the most structurally important news of the week. It comes in a context of extraordinary growth — $65 billion raised in Series H, valuation near a trillion — and signals that major frontier AI companies are starting to seriously consider public markets. Keeping PBC (Public Benefit Corporation) status is a strong commitment: Anthropic is seeking to prove that going public is compatible with a long-term safety mission. This precedent will have implications for OpenAI and other players that may follow this path.

The Physical AI wave — toward AI that acts in the real world. June 1 marks a remarkable convergence around physical AI: NVIDIA Cosmos 3 (the first open-weight omni-model for robots and agents), Cosmos Coalition (an open Runway+NVIDIA ecosystem), Luma OPAL Lab (open research on robotic generalization), MiniMax M3 (1M context + native multimodality), Qwen-VLA (11 types of robots under a unified architecture), and Qwen3.7-Plus (GUI+CLI agent in action in interfaces). This cluster is no coincidence — it reflects a growing consensus that AI’s next frontier is action in physical space, and that open source has a central role to play there. Competition is intensifying between closed and open models, especially in robotics and computer use benchmarks.

AI governance and safety are entering positive law. OpenAI’s publication of the Frontier Governance Framework, aligned with California’s TFAIA and the EU AI Act, and of the third-party evaluations guide marks a step forward: major AI companies are no longer defining their rules purely internally. They are aligning them with concrete legislation and publishing their evaluation methodologies. The Rosalind Biodefense initiative illustrates the other side of this governance — putting frontier AI at the service of social defense (pandemic preparedness, biosafety), with appropriate controls.

Developer tooling is accelerating and fragmenting. Claude Code receives five releases in four days, including Auto Mode on Bedrock/Vertex/Foundry and automatic loading of local plugins. Gemini CLI v0.44.0 unifies its Auto mode and expands IDE integrations. GitHub Copilot is refining its adoption taxonomy (4 phases, 28-day window) to help organizations measure the real maturity of their AI adoption. Developers now have access to dense, rapidly evolving agentic tooling — the question is no longer access to models, but whether teams can integrate these tools into their workflows in a coherent and measurable way.