⏰ Limited-Time Free Models

Several capable models are nearing the end of their free tier availability on OpenRouter. Three models expire tomorrow, June 5: Qwen3 30B A3B (131K context, $0.09/$0.45 per M tokens), Llama 3 Euryale 70B v2.1 ($1.48/M tokens), and Hermes 2 Pro — Llama-3 8B ($0.14/M tokens). The DeepSeek V3.1 Nex N1 ($0.14/$0.5 per M tokens, 131K context) runs until June 8, while Claude Opus 4.6 Fast ($30/$150 per M tokens, 1M context) remains free until June 29.

🏛️ DeepSeek Nears $7.4B in First-Ever Funding Round

DeepSeek is closing in on a record-breaking $7.4 billion funding round — its first ever — at a $59 billion valuation, Bloomberg, CNBC, and Reuters report. The round is reportedly led by Tencent, marking a pivotal moment for the Chinese AI startup that has shaken up the global AI pricing landscape. DeepSeek’s V4-Pro model recently received a permanent 75% price cut, and the company is previewing its next-generation V4 model running on Huawei Ascend hardware. This funding positions DeepSeek as a serious competitor to OpenAI and Anthropic on the global stage.

🤖 Anthropic Ships Claude Opus 4.8 with 3× Cheaper Fast Mode

Anthropic has released Claude Opus 4.8, introducing effort controls, dynamic workflows, and a 3× cheaper fast mode, per VentureBeat and The New Stack. The model also features improved honesty and reduced deception, with near-Mythos level alignment. Workflows are now capped at 1,000 subagents. Separately, Opus 4.6 Fast mode entered preview for GitHub Copilot. The release comes as Anthropic has submitted a confidential S-1 filing, eyeing an October IPO at a near-trillion-dollar valuation.

🏛️ Trump Signs AI Executive Order for Voluntary Federal Model Vetting

President Trump has signed an executive order establishing a voluntary framework for federal review of frontier AI models for national security risks, NPR, CNBC, and the NYT report. The order asks AI companies to voluntarily submit top models for cybersecurity testing, giving the government 30 days of early access. The framework includes a classified benchmark to determine which models qualify. OpenAI has reportedly diverged from the White House on specific safety provisions, per Politico, signaling continued tension between the administration and frontier labs over the right approach to AI governance.

🧠 Anthropic & DeepMind Now Actively Investigating AI Consciousness

Leading AI labs are expanding research into machine consciousness. The Financial Times reports that Google DeepMind, Anthropic, and Meta are actively investigating whether advanced AI models may be conscious or could become so. Richard Dawkins has publicly concluded that AI is already conscious, even if it doesn’t know it, while Anthropic’s CEO acknowledged in the NYT that “we don’t know if the models are conscious.” Separately, the Atlantic published a rebuttal arguing AI is not conscious — underscoring that even the framing of the debate itself is a sign of how far the technology has come.

🤖 Microsoft Build 2026: Agent-First Era with 7 Self-Developed Models

At Microsoft Build 2026, the company announced seven self-developed AI models (the MAI series) alongside Project Solara and the Majorana 2 quantum processor, per multiple reports. Microsoft claims its latest models beat Claude and Google’s models on key benchmarks. The company also introduced Frontier Tuning, a technique to teach AI how enterprises work at the organizational level — not just context injection. Microsoft Scout, an always-on personal agent, and expanded Copilot capabilities were also unveiled, cementing Microsoft’s “Agent-first” strategy.

In Brief

  • OpenAI diverges from White House on AI safety rules — Politico reports OpenAI is pushing back against specific safety provisions in the administration’s voluntary framework, creating a rare public rift.
  • Only 11% of production agents pass the AI agent security bar — A sobering finding from Help Net Security and Microsoft, revealing that the vast majority of deployed AI agents fail basic security benchmarks.
  • Perplexity CEO: economic efficiency decides the AI race — Aravind Srinivas told CNBC that cost per inference, not raw capability, will determine who wins; the company also unveiled its Perplexity Computer.
  • Google reportedly buying Android app code for AI training — 9to5Google reports Google is acquiring Android app source code from Play Store developers to train its AI models, raising privacy and consent questions.
  • AI chipflation spreading from data centers to the wider economy — Morgan Stanley warns that rising AI chip costs are beginning to affect sectors beyond tech.
  • Morgan Stanley opens trillion-dollar wealth management to AI agents — The bank will allow AI agents to operate within its wealth management funnel, a landmark moment for enterprise AI adoption.
  • Alibaba launches Qwen3.7-Plus — Alibaba’s Qwen team unveiled the multimodal agent model on the Bailian platform, supporting text, video, and imagery inputs at $0.4/$1.6 per 1M tokens.

🤗 Hugging Face Ecosystem

Hugging Face saw relatively quiet activity today. The most notable upload is OpenTransformer/AGILLM-4, a PyTorch long-context transformer with 7 likes, and a handful of experimental Qwen3-based checkpoints from individual researchers. Most uploads registered zero downloads — the platform is between major release cycles.

⭐ GitHub Trending: AI Edition

pewdiepie-archdaemon/odysseus continues its meteoric rise, now at 42,844 stars — a free, self-hosted AI workspace in JavaScript that has seen explosive growth since its launch. asz798838958/aBaiAutoplus (1,375★) is a Python toolkit for multi-platform AI account automation and ChatGPT Plus activation. ClaudioDrews/memory-os (743★, Python) provides a 7-layer memory operating system for Hermes Agent with Qdrant-backed persistent memory. 2aronS/Duel-Agents (700★, TypeScript) is a CLI, SDK, and IDE plugin ecosystem for agent-based development. liyue-aigc/female-portrait-director (503★) is a modular Codex Skill for directing AI portrait prompts.

🚀 New on OpenRouter

Qwen3.7-Plus has arrived on OpenRouter — Alibaba’s multimodal agent model with 1M context, priced at $0.4/M tokens (prompt). Limited-time free models (see ⏰ section above) also remain available for testing.

💡 Key Trends

Today’s AI landscape is defined by three converging forces: the funding super-cycle (DeepSeek $7.4B, Anthropic IPO near-trillion valuation), the governance pivot (Trump’s voluntary EO framework, OpenAI’s divergence, state-level lawsuits), and the agent security awakening (11% pass rate, Microsoft’s Frontier Tuning, Morgan Stanley’s agent deployment). The emergence of AI consciousness research as a mainstream investigation by top labs signals that the frontier is expanding beyond pure capability into the philosophical and regulatory dimensions of what these systems are becoming.