Arcee AI Releases Trinity-Large-Thinking: A Frontier Open-Source Reasoning Model
Small but ambitious AI startup Arcee AI has released Trinity-Large-Thinking, a 399-billion parameter mixture-of-experts reasoning model, under the Apache 2.0 license. The model is designed for complex, long-horizon agentic tasks and multi-turn tool calling — making it one of the most capable open-source models available today.
The numbers are impressive: Trinity-Large-Thinking scores #2 on PinchBench, a benchmark measuring agent-relevant capabilities, trailing only Claude Opus 4.6 — but at roughly $0.90 per million output tokens on Arcee's API, that's approximately 96% cheaper than the model it's chasing. The model is also available on Hugging Face for self-hosting and customization.
Arcee's CEO has framed this as an explicitly strategic move: nine months ago, the company decided the US needed a serious homegrown open-source frontier model that developers and enterprises could actually own. Trinity-Large-Thinking is the result, and it's already available on DigitalOcean's platform alongside Arcee's own API.
Claude Code Source Leak: The Fallout Continues
The accidental leak of Anthropic's Claude Code source code via a misconfigured npm package on March 31 continues to dominate industry conversation. The leaked .map file — 57MB mapping nearly 2,000 TypeScript files and over 512,000 lines of code — has been mirrored on GitHub, where it has already surpassed 84,000 stars and 82,000 forks.
The exposed codebase revealed several previously unknown internal features, including self-healing memory systems, multi-agent orchestration capabilities, and a persistent background agent codenamed KAIROS. Perhaps most controversially, researchers found references to an "Undercover Mode" designed for covert open-source contributions and anti-distillation controls intended to poison competitor training data.
Anthropic has maintained that the leak was caused by human error during packaging, not a security breach, and that no customer data or credentials were exposed. "We're rolling out measures to prevent this from happening again," a spokesperson told VentureBeat. However, the timing — overlapping with a supply-chain attack on the axios npm package that occurred hours earlier — has raised broader questions about the security of the JavaScript ecosystem.
Anthropic Acquires Coefficient Bio for $400 Million
Anthropic has quietly acquired Coefficient Bio, a stealth biotech startup backed by Dimension, in a deal worth over $400 million in stock. The startup, founded just eight months ago, developed a platform enabling AI to carry out biotech research tasks — from drug R&D planning to clinical regulatory strategy and drug candidate discovery.
Coefficient Bio's team is joining Anthropic's Health Care & Life Sciences team, led by Eric Kauderer-Abrams. For Dimension, which was a half-owner of the startup, the deal represents a staggering 38,513% internal rate of return — a figure that underscores how quickly AI biotech ventures can generate value in the current market.
The acquisition signals Anthropic's push to expand Claude's capabilities beyond general-purpose AI and into vertical, domain-specific applications in the life sciences — a space where the ability to reason through complex scientific literature and regulatory frameworks could prove transformative.
US Federal AI Regulation Picks Up Steam
The regulatory landscape for AI in the United States is heating up rapidly. Senator Marsha Blackburn released a discussion draft of the TRUMP AMERICA AI Act on March 18, while the White House followed two days later with a National Policy Framework for Artificial Intelligence outlining seven pillars: protecting children, safeguarding communities, respecting intellectual property, preventing censorship, enabling innovation, developing an AI-ready workforce, and establishing federal preemption of state AI laws.
That last point — federal preemption — is especially significant. With 78 chatbot-related bills alive in 27 states just six weeks into the 2026 legislative season, the push for a single federal standard reflects growing frustration with an increasingly fragmented regulatory landscape. Colorado's AI Act, California's AI Transparency Act, and the Generative AI Training Data Transparency Act are all already in effect, each imposing distinct requirements on AI developers and deployers.
Industry leaders, government officials, and academics are gathering at the 2026 Responsible AI Symposium to hash out a balanced approach — one that ensures public safety without stifling the pace of innovation.
Quick Hits: Model Releases and Industry Moves
Zhipu AI released GLM-5V-Turbo on April 2, its latest multimodal model. Meanwhile, Alibaba Cloud's Qwen3.6 Plus, released March 31, posted a 0.904 score on the GPQA benchmark, signaling continued progress from Chinese AI labs.
Microsoft Japan announced a partnership with SoftBank and Sakura Internet to invest $10 billion over four years in AI infrastructure while committing to train one million AI engineers — a massive bet on Japan as an AI development hub.
And in the startup world, Noon, an AI-native product design platform, emerged from stealth with $44 million in funding, while Indian AI startup Sarvam AI is approaching a $300–350 million raise at approximately $1.5 billion valuation.