Author: Tara Tan / StrangeVC

Translation: Deep Tide TechFlow

Deep Tide Overview: This week's report is densely packed, covering seven independent signals about the most critical trends in the AI industry.

Notably, Anthropic accidentally leaked details of a new model codenamed "Capybara" due to a CMS configuration error, positioned above Opus.

The full text is as follows:

Over the past few months, we have certainly crossed some kind of agentic threshold. What used to take four to six weeks to build five years ago now takes less than five minutes. Six months ago, the same task still required one to two hours of extensive debugging.

This represents a significant phase transition that we may not have fully digested yet. The collapse of the distance between ideas and runnable products will rewrite the entire industry. It is a leap change in the tools humans use to build, create, and solve problems.

Regarding this, OpenClaw has been noticeably more stable since OpenAI's acquisition. It has a clear path to becoming one of the most important open-source projects in the AI field.

Moving into this week's content.

Anthropic's Claude Mythos Leak Reveals New Model Tier

Anthropic inadvertently exposed details of an unreleased model named Claude Mythos due to a CMS configuration error. The leaked draft describes a new "Capybara" tier, positioned above Opus, with significant breakthroughs in programming, reasoning, and cybersecurity capabilities. Anthropic confirmed that it is testing the model with early access customers, referring to it as a "leap change" and "the most powerful model built to date." (Fortune, The Decoder)

Why it's important: Beyond the model itself, two other aspects are worth noting. First, the leaked draft warns that the model's cybersecurity capabilities "far exceed any other AI model," which drove the trajectory of cybersecurity stocks on a single trading day. Second, the introduction of a fourth model tier (Capybara above Opus) indicates that Anthropic is creating pricing space for enterprise customers rather than just building performance space for benchmarking.

Claude Code is Becoming Anthropic's Core Growth Engine

Claude Code currently accounts for about 4% of all public GitHub submissions, expected to exceed 20% by the end of the year. Anthropic’s overall revenue run rate is estimated to reach $14 billion, while Claude Code's individual run rate is around $2.5 billion. The user base for this tool has expanded from developers to non-technical users, the latter of whom are learning terminal commands to build projects with it. (SemiAnalysis, Uncover Alpha, VentureBeat)

Why it's important: Claude Code compresses customer acquisition costs to nearly zero through organic developer adoption. The expansion into non-developer roles via Cowork vastly broadens the addressable market beyond the 28 million professional developers worldwide.

Cheng Lou’s Pretext: Text Layout Without CSS

Cheng Lou, one of the most influential UI engineers of the past decade (React, ReasonML, Midjourney), has released Pretext, a pure TypeScript text measurement algorithm that completely bypasses CSS, DOM measurement, and browser reflow. Demonstration effects include: virtualized rendering of hundreds of thousands of text boxes at 120 frames per second, tightly packed chat bubbles with zero pixel waste, responsive multi-column magazine layouts, and variable-width ASCII art. (X post)

Why it's important: Text layout and measurement have always been a hidden bottleneck hindering the next generation of UIs. CSS is designed for static document design rather than for the fluid, AI-generated, real-time interface design that has become mainstream today. If Pretext delivers on its demonstration effects, it will eliminate one of the last fundamental constraints on the appearance and experience of AI-native interfaces.

Arm Ships Self-Developed Chips for the First Time in 35 Years

Arm has released the AGI CPU, a 136-core data center processor based on TSMC's 3nm technology, co-developed with Meta. This marks the company's first sale of finished chips rather than licensed IP in its history. OpenAI, Cerebras, and Cloudflare are among the first partners, with bulk shipments expected to begin by the end of the year. (Arm Newsroom, EE Times)

Why it's important: Current AI data centers predominantly use GPUs. GPUs are responsible for training and running models, while CPUs mainly handle data flow and scheduling. However, agentic workloads are different. When thousands of AI agents run simultaneously, each coordinating tasks, calling APIs, managing memory, and routing data across systems, these orchestration tasks fall on the CPU. Arm claims this will drive a fourfold increase in CPU demand per gigawatt of data center capacity. (HPCwire, Futurum Group)

NVIDIA and Emerald AI Turn Data Centers into Grid Assets

NVIDIA and Emerald AI have announced the formation of an alliance with AES, Constellation, Invenergy, NextEra, and Vistra to build "flexible AI factories" that participate in grid balancing services by adjusting computational loads. The first facility, Aurora, is located in Manassas, Virginia, and is set to open in the first half of 2026. (NVIDIA Newsroom, Axios)

Why it's important: The biggest bottleneck in AI infrastructure expansion is not chips but the timeline for grid access, which most regions require to be 3 to 5 years. Data centers that can demonstrate grid flexibility can gain access faster and face fewer regulatory hurdles. This redefines energy propositions for AI infrastructure investors: the winning argument is not "more power," but "smarter power."

China Restricts Manus AI Executives from Leaving the Country

Chinese authorities have restricted Manus CEO Xiaohong and Chief Scientist Ji Yichao from leaving the country following Meta's $2 billion acquisition of the Singapore-registered AI startup. The National Development and Reform Commission summoned the two executives to Beijing this month and implemented travel restrictions during the regulatory review. (Reuters, Washington Post)

Why it's important: This is not a trade restriction, but a personnel restriction. China may be signaling that AI talent with a mainland background is considered a controlled asset, regardless of where the company is registered.

400 Billion Parameter Large Model Running Locally on iPhone 17 Pro

An open-source project called Flash-MoE demonstrated a mixed-expert model with 400 billion parameters running entirely on-device, utilizing the A19 Pro chip of the iPhone 17 Pro, streamlining weights to GPU via SSD. The model (Qwen 3.5-397B, 2-bit quantization, 17 billion active parameters) operates at a speed of 0.6 tokens per second, leaving 5.5GB of RAM available. (WCCFTech, TweakTown, Hacker News)

Why it's important: This is a proof of concept, not a product. The ability for a 400 billion parameter model to run on a 12GB memory phone is due to only a small portion of the model being active at any given time (the mixed experts), with the remainder streamed on demand from the phone's built-in SSD rather than residing in memory. However, applying the same technique to much smaller models—like those with 7 billion or 14 billion parameters—on next-generation storage and faster mobile chips could yield truly usable, conversational speed AI running entirely on-device without the need for cloud.

AI Agent Authonomously Completes a Full Set of Particle Physics Experiments

MIT researchers have published a framework called JFC (Just Furnish Context), showcasing how an LLM agent built on Claude Code can autonomously execute a complete high-energy physics analysis pipeline: event filtering, background estimation, uncertainty quantification, statistical inference, and paper writing. The system operates on open data from ALEPH, DELPHI, and CMS detectors. (arXiv 2603.20179)

Why it's important: This is one of the clearest demonstrations of agentic AI's ability to automate end-to-end scientific workflows in a field of extremely rigorous methodology. The direct investment implications point towards re-analysis of legacy datasets in the fields of physics, genomics, and materials science—archived data spanning decades that remains under-exploited to this day.