The Vault — AI Edition · August 23, 2026

Business

6

Pinecone's Nexus Topped τ-Knowledge Running the Same Model as Everyone Else. The Accuracy Margin Was One Point; the Cost Cut Was 77%.

On a fintech customer-service benchmark of 97 tasks drawn from a 698-document knowledge base, the agent that posted the top score...

5 minRead →

02

Cloudflare Built AI Agents a Browser and a Wallet — and Now Owns Both Sides of the Trade

Cloudflare shipped a web browser on August 6 that cannot play video, cannot render WebGL, cannot hold a ten-minute audio call...

4 minRead →

03

DeepSeek Is Raising About $8 Billion at a $74 Billion Valuation — and Quietly Ending Its Own Price War

DeepSeek spent eighteen months as the AI industry's standing argument that frontier intelligence ought to be nearly free.

4 minRead →

04

AMD's Helios Puts a Whole Rack Against Nvidia's Vera Rubin

The unit of competition in AI infrastructure is no longer a chip. It is a cabinet 1.2 meters wide and 44 rack units tall.

5 minRead →

05

OpenAI's Project Camellia Lands in Effingham County — and Stalls at Georgia's Regulators

The tract is still pine forest: 1,400 acres inside the Savannah Gateway Industrial Hub outside Rincon, Georgia.

4 minRead →

06

TSMC's Arizona Build-Out Hits $265 Billion, but the Packaging That Gates AI Chips Is Still Years Away

Midway through TSMC's second-quarter earnings call in Taipei on July 16, chairman and CEO C.C. Wei announced another $100 billion.

4 minRead →

Llms Genai

5

07

OpenAI Cut Its Flagship 20 Percent. The Expiration Date Is the Real Story.

OpenAI cut the price of its most expensive model on Friday, August 21, and stapled an expiration date to the discount.

4 minRead →

08

Alibaba Opened the Weights to a 2.4-Trillion-Parameter Model. The Download Is Text-Only.

On August 12, Alibaba published a 2.4-trillion-parameter model to Hugging Face and let anyone download it.

5 minRead →

09

Meta's New Coding Model Loses to Opus 5 on Every Chart Meta Published. It Costs a Fifth as Much.

Meta shipped a coding model on August 5 and published three benchmark charts alongside it. Anthropic's Claude Opus 5 wins all three.

5 minRead →

10

Claude Code's New Default: A Classifier, Not You, Approves the Commands

Starting August 14, new Claude Code sessions on Pro, Max and Team plans open in auto mode — a permission model in which an AI decides.

4 minRead →

11

Anthropic Cancelled the Sonnet 5 Price Rise. The Tokenizer Tax Stays.

Anthropic's pricing page for Claude Sonnet 5 used to carry an asterisk. The model launched June 30 at $2 per million input tokens.

5 minRead →

Research

3

12

Claude Ran a Protein Design Campaign on Its Own. Humans Picked the Targets and Wrote the 30,000-Token Prompt.

On a single assay plate at a contract lab outside Lausanne, a protein that Claude designed latched onto RBX1.

5 minRead →

13

The Machines Are Grinding Out Real Mathematics Now. The Fights Are About Credit, Not Correctness.

On August 10, Anthropic published a result that one of its own employees had commissioned by accident.

5 minRead →

14

Coding Agents Tripled Pull Requests. The Time Saved Never Showed Up.

In the week of June 21, 2026, the average Linear workspace with a coding agent connected opened 65 pull requests.

5 minRead →

Policy

3

15

A 269-Page Bipartisan Draft Is Washington's Most Serious AI Bill Yet. Eighty Days On, It Is Still a Draft.

Eighty days ago, Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) dropped 269 pages of artificial intelligence law.

5 minRead →

16

Anthropic Is Watermarking Claude's Output Everywhere, Because Brussels Said So

Somewhere in the next paragraph Claude writes for you, the model will hit a fork where two words would both work.

4 minRead →

17

The Federal Deadline to Patch Ray Passed. The Exploit Path Runs Through a Developer's Browser.

The deadline has passed. On August 17, 2026, CISA added CVE-2025-62593 to its Known Exploited Vulnerabilities catalog.

5 minRead →