Anthropic shows managers what Claude spend buys

What happened: Anthropic now gives managers reports that show what their team's Claude use produced, what it cost, and where workflows broke down.

• Reports combine internal analytics with employee surveys, with a chosen date range and suggested fixes.
• A Jellyfish survey of 276,000 customers, cited by TechCrunch, found Claude Code and other Anthropic-based tools responsible for nearly a third of proposed code changes submitted for review.
• TechCrunch reports investor financials showed an operating profit in the second quarter and 900% net revenue retention.

Why it matters: Enterprise buyers keep asking AI vendors to prove a return on token spending, and Anthropic is answering with its own measurement layer. If you ship AI features, plan team-level cost and output tracking now, because buyers will ask for it during procurement.

What happened: Google's new flagship Gemini model, Carbon, is going to a first group of cyber defenders, with wider access promised later and no date given.

• Built for long engineering jobs that can run for days, such as tracing and patching bugs across large codebases.
• Cut successful cheating on Terminal-Bench 4.0 to 0.0 percent and unsupported claims by about half compared with predecessor Argon.
• Resisted every attack in an indirect prompt injection benchmark, with a 26 percent drop in overreach and a 29 percent drop in under-service on multi-step plans.

Why it matters: For founders building agents, the reliability and injection resistance numbers are the signals to watch. Wait to see how Carbon's pricing and access terms arrive before you change any product plans around it.

What happened: Federal loans of $5B will support eight power plants and add 1.1 gigawatts, mostly by raising output at existing reactors.

• The added output comes from higher-enriched fuel and equipment upgrades at existing reactors.
• Constellation already committed more than $7 billion to add 1,080 megawatts for Google, Amazon and Meta, and expects lower borrowing costs.
• The loan commitment is not finalized, and seven of the eight facilities are tied to those hyperscale buyers.

Why it matters: Electricity access and long-term power contracts are becoming the limit on how much compute you can deploy. Build capacity plans around grid timing and power cost, not only accelerator availability.