Twenty-dollar seats, four-hundred-dollar bills

A pricing-vs-burn post puts numbers on what most teams already half-know: a $20 Claude Pro or ChatGPT Plus seat used moderately is burning $200–$400/mo at API rates. GitHub Copilot is reportedly losing $20+/user/mo on a $10 plan; Anthropic users consume around $8 of compute per $1 of subscription. GitHub's own move to usage-based AI Credits on June 1 is the load-bearing tell.

Apple Silicon costs more than OpenRouter

William Angel

On a $4,299 M5 Max MacBook Pro running Gemma 4 (31 billion parameters), hardware depreciation — not electricity — sets the floor. The pessimistic case lands at $1.50–$4.79 per million tokens; OpenRouter serves the same model at $0.38–$0.50, with two-to-seven times the throughput.

“For apple silicon, the hardware cost dominates.”

Microsoft AI chief gives it 18 months

Fortune

Mustafa Suleyman tells the FT that most desk-bound work — accounting, legal, marketing, project management — will be fully automated in 12–18 months. Fortune's own piece carries the pushback in the same article: a 2025 Thomson Reuters report on professional services found marginal productivity gains, and a METR study found AI made software tasks take 20% longer.

I don't think AI will make your processes go faster

Frederick Vanbrabant

A Toyota Way / Goldratt re-read aimed at teams modeling AI as a throughput multiplier. Software is rarely slow at the keyboard; it's slow at scoping. Generated code doesn't collapse the scoping phase — if anything the documentation phase grows, because the agent needs every detail spelled out.

“Bottlenecks should receive predictable, high-quality inputs.”

Beyond code coverage: functionality testing with Playwright

Marlene Mhangami, GitHub

GitHub's COO puts the platform at roughly 275 million commits a week, with an extrapolation to ~14B by year-end. A Stanford study of 120,000 developers found unchecked AI lifted PR throughput while effective output rose ~1%, with refactor and rework eating the gains. The pitch: a Playwright MCP loop where the agent writes failing end-to-end tests against expected user behavior first.

“Clean code bases amplify AI gains; unchecked AI in a codebase is going to amplify entropy.”

Native all the way, until you need text

Artem Loenko

A 20-year Apple-platform developer walks through trying to ship a streaming Markdown chat UI in pure SwiftUI and fails: you can't select a Markdown document built from SwiftUI primitives, by design. TextKit 2 reaches parity at the cost of months of work; WebKit and Electron get there out of the box. Why most model clients end up on a browser engine.

“If you want to build rich text rendering for long-form chats, SwiftUI and Apple's native SDKs are not helping you. They stop being an advantage and start becoming constraints.”

The MCP hello page

Luke Lanchester

Customers were opening his mcp.acme.com/mcp URL in a browser, getting a 401 JSON blob, and filing tickets that said the link was broken. The fix: when the request is a GET with an HTML Accept header, return a plain HTML page explaining what an MCP server is and how to wire one up. Ticket volume dropped.

“It's not working though because they need to paste it into their client of choice, but no-one thinks that far ahead.”