|
AI Builders Digest
Saturday, September 26, 2026
|
|
The market share data Guillermo Rauch dropped this week is the most honest look at which AI labs are actually winning enterprise spend, and the story it tells is messier than any lab's PR team would like. Meanwhile, the rest of today's digest is a grab bag of practical signals and Google experiments, with varying degrees of substance.
|
|
---
|
|
01
|
Anthropic is losing ground fast, and Chinese models are eating the difference
|
|
|
Vercel CEO Guillermo Rauch shared real spend data from Vercel's AI Gateway covering the last two months, and it's striking. Anthropic dropped from 69% to 40% of total spend. OpenAI climbed from 10% to 24%. Kimi K3 and DeepSeek together absorbed roughly half of what Anthropic lost. Opus 5.5, which only launched two days ago, already accounts for 10% of spend on its own. OpenAI now leads in raw token volume, and handles 62% of image generation.
|
Why it matters: This is actual money moving, not survey data about developer sentiment. If your company is evaluating AI providers for a new project right now, the competitive landscape looks nothing like it did at the start of this year. Chinese models going from negligible to meaningful on a major cloud gateway also changes the procurement conversation: your legal team is going to have opinions about Kimi and DeepSeek that they didn't need to have six months ago.
|
|
Source →
|
|
---
|
|
02
|
The right way to prompt an AI agent: give it a number to hit
|
|
|
Peter Steinberger, a developer with a reputation for sharp AI workflow takes, posted a tip that's deceptively practical: vague instructions make agents stop too early, but giving them a quantified target forces them to actually work. His example: instead of "clean up the tests," try "remove 20% of the least useful tests while keeping code coverage within 2%."
|
Why it matters: If your team is using AI agents for codebase maintenance and getting underwhelming results, the problem is probably the brief, not the model. Agents optimize for task completion, and "clean up" completes the moment anything gets cleaned. A number gives the agent a harder target to satisfy and produces better output. This is the kind of prompt engineering that doesn't get blog posts but saves hours.
|
|
Source →
|
|
---
|
|
03
|
Google Labs is running an experiment about touching grass
|
|
|
Josh Woodward, who leads Google Labs, posted about Dreambeans, a new experiment that gives users a limited daily allotment of "beans" each morning, which point them toward real-world activities with real people. The concept is intentionally low-tech and offline-oriented.
|
Why it matters: This is a curiosity more than a product signal. A Google Labs team building something designed to get you off your phone is either a genuine values experiment or an elaborate way to seem relatable during an antitrust cycle. Worth watching to see if it goes anywhere.
|
|
Source →
|
|
---
|
|
Two items in today's feed don't clear the bar for a full write-up. Swyx posted a brief comment about conference scheduling without enough context to make it actionable. Thariq shared a link to a piece about AI and creativity with a single sentence of commentary. Both are link shares that don't stand alone. If the underlying articles are worth your time, you'll find them through the links above.
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|