US firms flee to Chinese models 🤖, Claude Code's origin story 🧑‍💻, price per token is a lie 💰

American companies are increasingly building on Chinese models like DeepSeek's and Z.ai's, which narrow the gap with frontier US systems while costing far less to run. CNBC reports the share of US tokens routed to Chinese models through OpenRouter has held above 30% every week since 8 February. The pull is cost: US labs keep raising prices, leaving teams with steep bills and hunting for cheaper options that hold up.

Top AI companies including OpenAI and Anthropic are showering startup founders with computing credits and competing discounts to win their business before rivals do. One founder building an AI-voice startup fielded escalating offers, and Cursor, the coding company now owned by SpaceX, ran a 75% discount through 5 July. For founders, the cost of building on frontier models is, for now, a bidding war in their favour.

Nvidia's Kyber rack, the cabinet built to run 144 of its 2027 Rubin Ultra chips as one machine, has slipped more than twelve months to 2028, according to SemiAnalysis, which blames trouble manufacturing a key circuit board. It is the latest in a run of setbacks questioning whether Nvidia's annual cadence is hitting hard manufacturing limits. The delay could hand AMD and Google a rare opening at the top of the market.

The New York Times reports that Alibaba's Qwen models have won over developers worldwide, yet the company struggles to turn that adoption into money because the models are open source and free to use and modify. It is the open-weight paradox in one company: the licence that drives distribution removes the obvious way to charge. For builders betting on free Chinese models, it raises how long their funding lasts.

A year ago tech leaders warned AI would wipe out jobs; over the past month the same chief executives have shifted to a far more optimistic tone about keeping people central to the work. Sam Altman, long a predictor of seismic workforce change, conceded the industry has been roughly right on technology and pretty wrong on the social and economic effects. The reversal signals how uncertain the picture still is.

Tencent's Hunyuan team made the full Hy3 public, a 295B-parameter mixture-of-experts model with 21B active, and in a reversal from April's preview put it under the Apache 2.0 licence. That licence shift is the real story: earlier terms excluded the EU, UK and South Korea, so legal teams killed deployments before evals finished. Reviewers say it beats Z.ai's GLM-5.2 at half the active size, trailing mainly on coding.

Anthropic's interpretability team offers evidence that Claude has a small collection of internal neural patterns that play a special role, set apart from all its other processing, which they call the J-space after the Jacobian method used to find them. Each pattern links to a word and lights up when that word is on the model's mind, not when it is said. They liken it to consciously accessible thought a person can describe and control.

Lilian Weng's 28-minute guide treats the harness, the scaffolding wrapped around a model, as the real engineering surface now that raw capability is commoditising. She lays out three design patterns, workflow automation, the file system as persistent memory, and sub-agents with backend jobs, then works a coding-agent case study. The rest covers optimising the harness: context engineering, workflow design, and self-improving harnesses that evolve their own scaffolding through evolutionary search.

Cutting through the noise about designing loops instead of prompting, the Claude Code team defines a loop plainly as an agent repeating cycles of work until a stop condition is met. They sort loops by how they are triggered, how they stop, which Claude Code primitive drives them, and which task each suits, starting with the turn-based loop every prompt already runs. The advice: reach for the simplest option and add heavier patterns selectively.

The VS Code team ran a two-week experiment with OpenAI to tune GPT-5.5's system prompt inside their coding harness, the layer connecting the model to tools, context and the agent loop. The question was simple: nudge the agent to explore less and validate sooner, and see if it gets faster and cheaper without getting worse. They tested two prompt changes against a control on live traffic and kept the winner.

Google's Pathways and MaxText team walk through elastic training on Cloud TPUs, deliberately failing a worker mid-run and watching the training process recover in place within seconds rather than relaunching the whole job from its last checkpoint. For anyone running large multi-node training, that removes the tax where one machine dying forces a restart and burns hours. It is a concrete introduction to wiring elastic training into a JAX and MaxText stack.

Martin Alderson argues the least-understood shift coming in AI economics is not training cost but inference margin. His napkin maths puts the gross margin on frontier inference near 90% when Anthropic or OpenAI charge $25 per million tokens, a spread held up by lack of competition, not real compute cost. Open-weight models reaching frontier quality, GLM 5.2 among them, are the arbitrage, and he calls this the real DeepSeek moment.

Jan Ilowski argues that comparing models by dollars per million tokens breaks two ways. Each lab tokenises differently, so the same text costs different counts, and Anthropic recently made Claude split text into 30% more tokens, a stealth price rise. Hidden reasoning tokens then bill at output rates but vary wildly. His cost-per-task table reorders things: GPT-5.5 finishes at nearly half of Opus 4.8's cost, and GLM-5.2's cheap tokens aren't proportionally cheap.

Anthropic's oral history traces Claude Code from 2021 coding-model research through clide, a wonky internal CLI, to Boris Cherny's two-day Claude CLI demo that drew three Slack likes and became the product. The team stayed deliberately small, pushed fixes within minutes, and rode the Claude 4 capability jump plus a subscription model into takeoff. By winter 2025 Cherny says every line of his code, 88 commits in one day, came from Claude Code.

In a SemiAnalysis conversation, the case is that robots are the real general-purpose technology because they decouple capital from human labour for the first time, and everyone keeps underestimating Unitree the way they did DJI and BYD. The hard part is not intelligence but hardware: China's vertical integration and actuator manufacturing make its humanoids far cheaper to build than American ones. The bottleneck is industrial, and no model closes it.

Deptrust is a CLI and MCP server that checks package versions for known vulnerabilities across npm, PyPI, crates.io, Go modules, Maven and a dozen more registries, calling the public OSV and registry APIs directly with no hosted service in between. It was born from AI agents constantly pulling in old, vulnerable versions, and returns a plain verdict of block, review or allow by highest severity. As an MCP server, your agent can check itself before shipping.

Screencommander is a macOS CLI and MCP server that gives agents an observe, decide, act loop over the desktop, reading the screen as structured accessibility-tree data when it can and grabbing pixels only when it must. Actions run through a tier ladder, from cursor-free accessibility calls to events posted straight to an app, so the pointer never moves and background windows work. Because it exposes the element tree as text, non-vision models can drive too.

Whim is a Go library and CLI that opens a throwaway root shell inside an AWS Lambda Firecracker microVM in roughly two seconds, runs a command or an interactive session, then lets the VM vanish. After a one-time image build, every launch is fresh and isolated, with a server-side time-to-live so nothing lingers if a client dies. It is a tidy primitive for running untrusted or agent-generated code in a disposable sandbox.

Cad-skill is a Claude Code skill that turns a described physical object into a parametric, 3D-printable model using CadQuery. Ask for a wall mount for an Arduino or a snap-fit lid and Claude picks up the skill from trigger keywords, gathers requirements, builds the model in phases, and hands back an STL plus a rendered preview. It runs on Python 3.10 to 3.12, and the write-up shows real parts published to MakerWorld.