At Connect, Zuckerberg made tokens nearly free and bet the company on skimming transactions instead. That quietly rewrites who pays for AI.
Tonight’s edition draws on four desks — X, Semafor, The Information and TechCrunch — and ranks each story by corroboration: how many of those desks independently carried it, with significance breaking ties. The day’s theme is state as the new lock-in: agents are acquiring email addresses, payment rights and cached context inside vendor platforms, just as the labs cut sticker prices and governments argue over who should hold the off-switch.
Meta used Wednesday’s Connect keynote in Menlo Park to declare Muse — its personal agent, only weeks old — the centerpiece of the company. Zuckerberg framed it as the future “personal superintelligence” for billions, and announced a slate of new surfaces: a real-time video avatar (powered by a new Muse Realtime Avatar model), wake-word integration on Meta’s smart glasses “in the coming months,” full computer-use on the Mac, and — tellingly — Muse getting its own email address so it can be cc’d on threads and work in the background.
The commercial architecture is the real news. Meta says it will make Muse “free for a huge number of tokens,” expecting to profit “by taking a small fee from transactions.” It has wired in Shopify’s full catalog and Shop Pay, PayPal and Stripe, plus retail connectors for Best Buy, Gap, Sephora, Walmart and Wayfair, travel via Expedia, and Instacart soon. On the work side, GitHub, Granola and Notion connectors ship now, and Meta says more than 1,500 developers applied to build connectors in under a week.
The competitive ripple is already visible. The Information reports OpenAI is developing features to counter xAI’s Grok bot and weighing a response to Muse (headline only; paywalled). The same day, OpenAI pushed voice-driven agentic tasks into the ChatGPT mobile Work tab. Earlier in the week TechCrunch reported Meta conceded Muse’s design was deliberately modeled on the open-source OpenClaw agent.
Why it matters to a CTO: an ad-funded giant pricing agent inference near zero changes the reference price your employees carry into work. More importantly, a transaction-fee model means Muse’s incentives are aligned with completing purchases, not with your procurement policy. If you sell online, agent-originated traffic arriving through Shopify, PayPal and Stripe rails becomes a channel with a platform tax — plan for “agent-legible” catalogs and APIs now. If you run IT, a consumer agent with Mac control and its own inbox is a shadow-IT vector; your DLP and identity policy should name it explicitly.
Sources: TechCrunch — Everything new coming to Muse · TechCrunch — Muse wearable · TechCrunch — OpenClaw likeness · The Information — OpenAI mulls response to Muse (headline only)
Everyone is reading this week’s pricing through the headline input/output rates. Look instead at the line most reports skip: Opus 5.5 cuts cache reads from $0.50 to $0.20 per million tokens. In agentic workloads, the same system prompt, tool schemas and working context are resent on every step, so the cached share of tokens routinely dwarfs fresh input. The effective bill is set by the cache rate — and a cache lives on one vendor’s servers.
That makes cheap cache reads a gravitational field. The more your agents run, the more warm state you accumulate with a single provider, and the more expensive it becomes to route a given workflow elsewhere (you pay full cold-start price to rebuild context). Meta’s Muse is the consumer version of the same move: free tokens, but the agent’s inbox, connectors and purchase history accrue inside Meta. And the UN’s CIOs, per Semafor, are drafting “exit plans” precisely because Mythos access was cut off once already.
The counter-intuitive conclusion: in a price war, the cheapest vendor is not the one with the lowest rate card — it is the one whose state you can walk away from. Measure it.
Avatar, glasses, Mac control, its own email address, and a commerce stack (Shopify, Shop Pay, PayPal, Stripe, major retailers). Tokens near-free; Meta takes a cut of transactions. OpenAI is reportedly weighing a response.
Anthropic’s Opus 5.5 lands at $4/$20 per million tokens with what TechCrunch calls “Fable-level performance”; OpenAI’s GPT-6 Sol ($2/$10) and Luna ($0.10/$0.50) roughly halve the prior tier. On X, even xAI’s Grok account named Opus 5.5 the top coding model. Adjacent: The Information reports software incumbents are discounting their own AI to keep customers from defecting to the labs.
Amodei told the council AI “could be a risk to humanity as a whole” and proposed narrow global bans (e.g., AI for bioweapons), verification regimes and a shared incident-notification system. Altman said decisions “cannot be made by labs in San Francisco alone.” US OSTP chief Michael Kratsios answered that Washington “totally rejects” international efforts at global AI governance. Trump hosts Xi today; analysts expect at most shared safety definitions and an incident hotline.
The ITU’s CIO Khuloud Odeh says last year’s ICC sanctions and this year’s Mythos model cutoff were a “wake-up call.” UN agencies are exploring shared compute and data infrastructure, sorting systems by sensitivity and asking “what would be an exit plan?” — with “sanctions-driven service suspension” now a named cloud risk.
Anthropic says its Bay Area lab surfaced a previously unknown phage enzyme system that can cut, copy and paste DNA — found “mostly, though not entirely, by Claude” in about 21 hours using ~950 agents and 210M tokens. Humans performed all bench work, at BSL-1/2 only; Amodei acknowledged a similar Stanford finding. Validation by the field is pending.
The Information’s Applied AI desk reports a single vulnerability class spanning all four major coding agents. Details are paywalled; we could not verify the specifics. Context: Cisco Talos this week disclosed CLOSEDQUORUM, an implant that lets a four-model LLM panel vote on its next post-exploitation step, with no human operator.
Plus and Pro users can now draft documents, summarize Slack, build sites and use the cloud browser by voice from their phones, with handoff to desktop. Free and Go users get plugins and connected apps.
An exclusive reports DeepSeek is committing heavily to Huawei silicon. Paywalled; headline only. It lands as Trump and Xi meet with chips and AI on the agenda.
Anthropic and Amodei lead with the find and the speed (21 hours, ~950 agents). TechCrunch frames the bigger reveal as the opposite — that humans, not Claude, still run the bench — and places it against the same CEO’s bioterror warnings and his call to “pace the frontier.” Read it as both: a capability signal and a template for human-gated autonomy.
At the UN, both CEOs urged international standards. Semafor’s earlier reporting shows OpenAI simultaneously pitching a US-led safety coalition, while the administration rejects international governance outright. The public “call for global oversight” framing overstates convergence: the live debate is over who leads, not whether to coordinate.