// daily-digest · mon 04 may 2026 · morning refresh · last 72h focus
Three AI-lab money-and-strategy stories collided over the weekend: Anthropic is fielding offers for a $50B round at a $900B valuation, OpenAI bought TBPN outright, and DeepSeek's V4 preview is now within a model generation of Gemini 3.1 Pro on world knowledge. Plus Felix Rieseberg on why Claude needs its own computer, Stripe's 288-launch firehose distilled, a Simon Willison sycophancy quote, Cat Wu on Anthropic's shipping cadence, a 100k-person GLP-1 mental-health study, and a wildcard from Bronze-Age Spain.
— refreshed for you, in 10 items.
Why for you: The number more than doubles February's $380B mark and would push Anthropic past OpenAI's $852B. Reported $30B annualized revenue. Board decides on the round in May. The Pentagon snub from last week reads very differently when the buyer line forms at $900B.
CNBC: at $900B Anthropic would surpass OpenAI as the most valuable AI startup
Why for you: Thompson on OpenAI buying a tech podcast network — "makes no sense, which may be par for the course." Pair with the token-tsunami framing of how compute spend is now reshaping every line on every hyperscaler's earnings call. Companion to last week's Anthropic-Google TPU piece.
Why for you: The official roundup of everything from Sessions 2026 — the Agentic Commerce Suite, 250M consumer wallets linked to AI agents, streaming payments on Tempo, the Machine Payments Protocol co-authored with Tempo, Stripe Projects GA with 14 new partners (Render, Twilio, Sentry, WorkOS, Browserbase, GitLab, ElevenLabs). One link to absorb the whole conference.
Stripe blog: the Agentic Commerce Suite — selling on AI agents end to end
Why for you: Direct relevance: this is the engineering thesis behind the product you're literally talking to right now. Rieseberg (ex-Slack/Electron) walks through why Cowork came out of an "accident" — people using Claude Code for knowledge work — and why local-first agents with their own VM beats cloud-only. Skills as markdown, real workflows: video, taxes, calendars.
Why for you: Tiny post, big idea: Anthropic built an automatic classifier that judges sycophancy by whether Claude pushes back when it should. The implication is that "willingness to disagree with the user" is now a measurable, optimizable model property. Worth keeping in your back pocket next time someone claims AI evals are unfalsifiable vibes.
Why for you: V4-Pro: 1.6T parameters, open-source, ~4x cheaper than US peers, beats every open competitor on math/code, trails only Gemini 3.1 Pro on world knowledge. Co-designed with Huawei Ascend 950 supernodes. The 3-to-6-month gap behind US frontier is the new center of gravity for the export-controls debate.
Why for you: Wu's claim: Anthropic's product cycles went from six months to one month — sometimes one week, sometimes one day. Two drivers: early access to unreleased models (Mythos), and a launch-room process that compresses spec-to-ship. Most useful PM tactic she names: ask the model to introspect on its own mistakes. Cat also runs Cowork — see [04] for the engineering side.
Why for you: Worth re-reading in light of this week's news. Opus 4.7 is the model the $900B valuation is being underwritten on, the model in the Pentagon dispute, and the one Wu's team uses to compress launch cycles. SWE-bench Pro 64.3%, "users report being able to hand off their hardest coding work with confidence." GitHub Copilot bumped its premium multiplier to 15x on Apr 30 — the market is voting.
Why for you: 95,490 Swedes with anxiety/depression, 20K+ on GLP-1s. During semaglutide use: 44% lower depression risk, 38% lower anxiety risk, 42% fewer psychiatric hospitalizations, 47% lower substance-use hospitalizations, lower suicidal-behavior risk. Massive, register-based, hard to dismiss. Shifts the GLP-1 conversation from "weight-loss side effect" to "potential class effect on mood/addiction."
SciTechDaily plain-English summary if you want the TL;DR first
Why for you (off your normal lanes): Lead-isotope work has long suggested the metal in Scandinavian Bronze Age artifacts came from somewhere in southwestern Iberia, but the actual mines hadn't been found. Gothenburg researchers just identified six in Extremadura — copper, lead, silver — with grooved stone axes still on site. The lead investigator estimates ~150 prehistoric mines in the region remain undocumented. A neat case of an old isotope chart finally getting its physical reference.