Anthropic in Early Talks to Lease ~$10B of Meta Compute
Anthropic is in early discussions to lease computing power from Meta in a deal reportedly worth about $10 billion over two years, per the New York Times. The proposed structure has Anthropic paying Meta in monthly installments, with an option for either side to exit early. Anthropic reportedly floated the arrangement back in June, and both companies declined to comment.
The driver is raw capacity. Anthropic has capped usage on its most advanced models — including Claude Fable — because it doesn’t have enough compute to run them unthrottled. The talks echo its May agreement to pay SpaceX $45B over three years, and for Meta they’re the clearest sign yet it wants into the cloud business, on top of a projected ~$145B capex year.
Claude for Government Desktop Enters Public Beta on FedRAMP High
Anthropic put Claude Code and Claude Cowork into public beta through Claude for Government Desktop, delivered inside a FedRAMP High authorized environment. It ships as a desktop app that deploys through standard agency MDM platforms — no separate cloud-provider contract required — with local conversation history, department-level admin, tamper-evident audit logs, and hard not-to-exceed spend caps.
The pitch to agencies is delegation with guardrails: Cowork works directly with files to draft memos, review RFPs, and build decks, while Claude Code modernizes the software behind public services. The rollout extends Claude across all three branches of the federal government, with program offices able to tie AI spend to appropriated funds.
Claude Fable 5 Lands on Max and Team Premium Plans
With the free-access window for subscribers closing, Anthropic is folding Claude Fable 5 — its top creative-writing model — into Max and Team Premium plans at 50% of standard usage limits. It’s the model’s move from a temporary promo into a permanent, paid home after weeks of extended free trials.
The half-rate cap is telling. It lines up directly with the compute squeeze behind the Meta talks: Fable 5 is expensive to serve, so paid access comes metered rather than unlimited. A brief July 17 glitch that wrongly charged some users credits before the promo deadline was refunded, with matching credits added.
Claude Code Adds Background /fork, an EndConversation Tool, and Progress Heartbeats
The latest Claude Code updates lean into long-running agent work. Background sessions for /fork plus smarter /resume and /background flows let you spin off and reattach to work without babysitting a terminal. A new EndConversation tool lets agents cleanly wrap up, and periodic progress heartbeats keep long tool calls from looking frozen.
There’s a security thread too: stronger safeguards around WebSearch, subagent spawns, and Bash execution, plus fixes to single-segment dir/** allow rules and overly permissive Windows PowerShell and Bash permission checks. Reliability got broad attention across web search, MCP, worktrees, telemetry, and remote sessions.
New Agent-Memory Beta Header Brings Stable Ordering
The Claude Developer Platform added an agent-memory-2026-07-22 beta header that changes memory-listing behavior to return results in a stable, server-defined order. For anyone building agents that persist and replay memory, deterministic ordering removes a subtle source of flakiness when diffing or paginating stored entries.
It arrives alongside the week’s broader platform housekeeping: API key expiration settings with pre-expiry email reminders, consolidated three-tier rate limits, and self-serve HIPAA readiness configuration for Enterprise and API orgs. Small pieces, but they add up to a platform that’s easier to govern at scale.
Claude Cowork Goes Cloud: Mobile, Web, and Background Tasks
Anthropic is moving Claude Cowork to the cloud, taking the agent beyond a single desktop. Sessions and files now follow you across devices, with mobile and web access, background work that runs when your machine is offline, scheduled tasks, shared chats and projects, and mobile approvals for actions in flight.
For builders, the interesting shift is architectural: Cowork stops being a local process and becomes a hosted agent you supervise from anywhere. That is the same pattern showing up in Claude Code’s background sessions — long-running, resumable work that no longer depends on one always-on terminal.
LTM and UST Sign On as the Partner Network Scales
The Claude Partner Network keeps compounding. LTM announced a partnership (July 13) to speed enterprise Claude adoption, pairing Claude, Claude Code, and Cowork with its implementation muscle to move clients from pilots to production. Days earlier, UST struck a strategic alliance (July 10) aimed at making Global 1000 firms “AI-native.”
The momentum is real: more than 40,000 firms have applied to the network and over 10,000 consultants hold a Claude certification, backed by Anthropic’s $100M initial commitment. This is the go-to-market layer an enterprise platform needs — the people who actually get Claude into production.
“Reflect” Gives Users a Dashboard for Their AI Habits
Anthropic’s Reflect feature is a built-in dashboard that lets you track and visualize how you use Claude — and your broader AI habits — over time. Instead of usage disappearing into scattered chats, Reflect surfaces patterns: what you lean on Claude for, how often, and where the time goes.
It’s a small but notable turn toward user-facing metering and self-awareness. As agents take on more autonomous work, giving people a clear picture of what they’ve delegated — and what it cost — becomes part of using these tools responsibly.
The Real Constraint Isn’t Models — It’s Compute
Read today together and one theme dominates: capacity. A ~$10B Meta lease on top of a $45B SpaceX deal, Fable 5 arriving on paid plans at half the usual limits, and a candid admission that the best models stay throttled for lack of processing power — these aren’t separate stories. They’re the same story told three ways. Anthropic has demand it literally cannot serve, and it’s spending tens of billions to close the gap.
That reframes everything else on the page. The government desktop beta, Cowork in the cloud, background agents in Claude Code — all of it multiplies the compute a single user consumes, because agents that run in the background and across devices burn cycles even when you’re not watching. The product roadmap and the infrastructure scramble are two ends of the same rope. The company that wins the agent era won’t just have the smartest model; it’ll have the electricity to run it. Watch whether the Meta talks actually close — early-stage discussions have a way of quietly evaporating, and Anthropic’s usage caps are riding on deals like this landing.