Artificial Intelligence

Anthropic launched Claude Opus 5 on July 24, 2026: a model that approaches Fable 5’s performance while costing half as much per task. We cover the benchmarks, real use cases, and how to start testing it today.

The NSF allocated $83 million to a data backbone for AI-powered science, a quieter but technically concrete announcement next to the Genesis Mission’s $5 billion. The investment aims to give unified access to scientific datasets for training and validating models reproducibly.

A historian at the University of California, Santa Cruz used Claude Code to build the Book Prize Index, a free database of 6,500 nonfiction books that won or were finalists for major English-language prizes. The tool uses semantic search to replicate the experience of browsing a well-curated library shelf, as a direct response to the flood of generic AI-generated content.

Gigatoken is a Rust tokenizer that reaches 24.53 GB/s in benchmarks, up to 989 times faster than HuggingFace Tokenizers. It supports more than 15 model families, from Llama to DeepSeek, and installs with a simple pip install.

A public ChatGPT thread exploring a possible counterexample to the Jacobian conjecture put one of algebra’s longest-standing open problems back in the spotlight. We cover what the conjecture states, what’s actually been proven, and how to verify it yourself with Python.

The White House launched Genesis Mission, an investment of over $5 billion to apply artificial intelligence to science within the Department of Energy. Microsoft has already committed $60 million as an infrastructure partner.

Ben Thompson argues in Stratechery that open-weight models like Kimi K3 aren’t free to serve: inference cost, not license price, is what defines margin in the AI industry. The debate reopens the question of who actually wins when a model competes against a rival that charges less per token but needs more tokens to respond.

Ben Thompson analyzes in Stratechery why open Chinese models like Kimi K3 aren’t free to serve. The essay distinguishes AI COGS, a variable cost, from the fixed investment in research and development, and explains why tokens aren’t a fungible commodity.

Kimi K3 charges $3 per million input tokens and $15 for output, but that doesn’t mean it’s free to serve. A Stratechery analysis separates fixed R&D from variable COGS and explains why price per token doesn’t capture the real cost of intelligence.

Chinese open models like Kimi K2.6, Kimi K3, and Qwen3.8 are advancing on OpenAI and Anthropic’s closed models. An a16z analysis cited by The Economist estimates that 80% of startups already use at least one Chinese model in production, as Washington tightens GPU export controls.