AI bills climb as benchmarks and admin controls multiply

Even as AI token prices have fallen sharply, overall spending is rising, and Anthropic is giving Claude Enterprise admins new usage and spend controls. Meanwhile, Meta claims parity with OpenAI's top model, and Cursor updated its evaluation suite.

Cheaper tokens, bigger AI bills

Falling unit prices have not made AI cheaper to run at work. One analysis argues that while the price of AI tokens has plummeted, the total cost of what it calls AI employees has surged. Its title asks whether AI employees may now cost more than humans. For teams running agents or heavy assistant workloads, the gap between per-token pricing and the monthly bill is the number that matters.

That is the gap Anthropic is targeting with new analytics and cost controls for Claude Enterprise, which give administrators more visibility into who is using the product and what it costs. If you own an internal AI rollout, usage reporting at the seat and team level may help you defend a budget, or show where it is leaking.

Competing claims about model performance

Meta says its new model, Watermelon, achieves benchmark results comparable to OpenAI's GPT-5.5. If that holds up outside Meta's chosen tests, teams would have another frontier-class option when picking a model. Benchmark parity is a narrow claim, though, and says little about cost, latency or behavior on a specific codebase.

Cursor's own CursorBench 3.1 evaluation suite offers a different reference point. It is an updated set of evaluations for assessing how models perform, from a company whose product depends on picking the right model for coding work. Comparing a lab's self-reported scores against evaluations closer to your actual tasks is a reasonable check before switching providers.

A shared unit of cloud compute

AWS's weekly roundup describes a session in which AWS, Google and Microsoft appeared together, and the companies reportedly agreed on a common unit of compute. Few details beyond that are available, so this is best treated as an early step rather than a finished standard. If it develops further, a shared definition could make capacity and pricing easier to compare across the three largest clouds, which matters for anyone running multi-cloud workloads or negotiating commitments.