Daily AI News - 2026-05-23

⏰ Limited-Time Free Models

Several models on OpenRouter are approaching the end of their free tier availability:

🚀 DeepSeek Permanently Slashes V4-Pro Price by 75%

China’s DeepSeek announced a permanent 75% price cut on its flagship V4-Pro model, in a move that Reuters and multiple outlets are calling a major escalation in the AI pricing war. The aggressive restructuring comes amid growing global demand for affordable frontier AI and follows the earlier V4 launch that shook markets last year.

• Pricing: The permanent reduction brings V4-Pro to a fraction of its original cost — a direct challenge to Western AI labs that have been raising prices.
• Market impact: The move is widely seen as DeepSeek weaponizing its cost-efficient training pipeline to pressure competitors. Android Headlines notes it targets “Western AI rate limit anger” as enterprises seek cheaper alternatives.
• Timing: The cut follows a NIST CAISI evaluation of the model and comes as DeepSeek increasingly receives state backing for its AI ambitions, per Fortune.

🏛️ Trump Cancels AI Executive Order at the Last Minute

President Trump abruptly cancelled the signing of a landmark AI executive order that would have granted the government broad oversight of AI model releases. Multiple reports indicate the decision came after heavy lobbying from Silicon Valley.

• The reversal: Trump told CNBC he “didn’t like certain aspects” of the order. The WSJ reports the cancellation was driven by concerns about overregulation.
• Silicon Valley influence: Politico reports that David Sacks, Trump’s AI and crypto czar, raised industry concerns directly. The Washington Post notes pressure from tech executives helped block the expected order.
• International context: Reuters reports the postponement was partly motivated by a desire to compete with China on AI capability rather than impose regulatory constraints.
• The cancelled order: Politico published the unsigned executive order text, which would have required government review of frontier AI models before public release.

This marks a major shift — the administration that previously signed AI deregulation executive orders now appears divided between industry growth and security concerns.

🔐 NSA Publishes MCP Security Guidance

The National Security Agency released official security design considerations for AI-driven automation using the Model Context Protocol (MCP) — directly relevant to the developer ecosystem around agentic AI.

• The guidance covers secure deployment patterns for MCP-based agent architectures, emphasizing zero-trust principles for AI tool access.
• The release comes alongside joint US-ally guidance on agentic AI system security, signaling growing government attention to the security implications of AI agents.
• For the Hermes Agent community, this represents official recognition of MCP as a critical infrastructure protocol worthy of security hardening.

🔍 Alibaba Qwen 3.7 Max: Autonomous 35-Hour Operation

New details have emerged about Alibaba’s Qwen 3.7 Max, the latest reasoning agent model now available on OpenRouter with a 1M-token context window:

• Autonomous capability: VentureBeat reports the model can run autonomously for 35 hours and supports external harnesses including Anthropic’s Claude Code — positioning it as a serious competitor in the agent space.
• Pricing: Available at $2.5/M tokens prompt on OpenRouter, maintaining Alibaba’s aggressive pricing strategy.
• Ecosystem: The model is the highest-ranking Chinese model on the Chatbot Arena leaderboard, and Alibaba is integrating it across Taobao (agentic shopping), automotive (voice assistants), and enterprise workflows.

🛡️ Microsoft Open-Sources RAMPART and Clarity for Agent Security

Microsoft released two open-source tools — RAMPART and Clarity — designed to bring safety into the AI agent development workflow. The Hacker News, CSOonline, and DevOps.com all covered the announcement.

• RAMPART: A red-teaming framework for AI agent systems, allowing developers to simulate adversarial scenarios and test agent security postures before deployment.
• Clarity: A tool for auditing and understanding agent decision-making processes, providing transparency into how autonomous agents reach conclusions.
• Significance: As agentic AI moves from research to production, security tooling like this becomes as essential as unit testing. Microsoft’s decision to open-source both tools reflects industry-wide recognition that agent safety is a shared responsibility.

💰 The “Tokenmaxxing” Cost Crisis Hits Big Tech

Tom’s Hardware reports that employees at Microsoft, Meta, and Amazon are exploiting internal AI platforms by inflating token usage — a phenomenon dubbed “tokenmaxxing” — leading to massive cost overruns.

• The problem: Agentic AI consumes up to 1,000x more tokens than standard chatbot interactions. Employees are gaming AI usage metrics to meet arbitrary targets or simply because AI tools are free internally.
• The response: Companies are pulling back on internal AI deployments, with some imposing hard token caps and monitoring systems.
• Broader implications: The Economist picks up the story, framing it as “the AI rush hitting a bottleneck” — the cost of running AI at scale is proving far higher than expected, and internal governance hasn’t caught up.

🤗 Trending on Hugging Face

New models spotted on Hugging Face this cycle:

• lllyasviel/In-Context-LoRA-Books — New LoRA-based in-context learning approach for image generation
• anziank/grio-qwen2.5-1.5b-coreml-anyLM-seq2048 — Qwen 2.5 optimized for CoreML/on-device iOS deployment (44 downloads, 1 like)
• Sachin21112004/distilbart-news-summarizer — News summarization fine-tune with 10 likes and growing downloads

⭐ GitHub Trending: AI Edition

• Doorman11991/smallcode ⭐ 1,302 — AI coding agent optimized for small LLMs, achieving 87% benchmark with a 4B-active model. (JavaScript)
• datawhalechina/Agent-Learning-Hub ⭐ 1,209 — Comprehensive AI Agent learning pathway and resource collection from DataWhale China. (HTML)
• lynote-ai/humanize-text ⭐ 545 — Open-source AI text humanizer that converts AI-generated content into undetectable human-like writing. (Python)
• LiuMengxuan04/shushu-internship-tool ⭐ 462 — Job-hunting AI tool: transforms job descriptions into projects and resumes into interview prep. (Python)
• basketikun/infinite-canvas ⭐ 454 — Open-source infinite canvas for AI creation with image generation, editing, prompt library and asset management. (TypeScript)

💡 Key Trends

1. AI regulation whiplash — Trump’s last-minute cancellation of the AI executive order shows the deep divide within the administration between pro-regulation security hawks and Silicon Valley-backed deregulation advocates. This mirrors the global pattern of stop-start AI governance.
2. Cost economics of AI agents — From DeepSeek’s 75% price cut to the “tokenmaxxing” crisis at big tech, the industry is grappling with the real economics of AI at scale. Low inference prices are a competitive weapon, but internal costs are proving harder to manage than expected.
3. Security goes mainstream for agentic AI — The NSA MCP guidance and Microsoft’s RAMPART/Clarity release both signal that AI agent security is evolving from an afterthought to a first-class engineering concern. Expect more government frameworks and open-source tooling in this space.
4. Chinese AI labs double down — DeepSeek’s price war and Alibaba’s 35-hour autonomous agent capability show Chinese AI companies are not just catching up — they’re defining new competitive dimensions in pricing and agent autonomy.