Anthropic · Claude · Cost efficiency
Claude got cheaper without becoming much dumber
Anthropic launched Claude Opus 5.5 on Tuesday. The company says it performs at the level of Fable 5.1 on most work and costs 40% less than Opus 5 on typical workloads. API prices are $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, while cheaper cache reads and lower token use account for the larger estimated workload reduction.
The exact benchmark rankings matter less than the economic direction. For much of the AI race, stronger models implied more compute and more money. Architecture, optimization, inference techniques and specialized hardware are weakening that relationship. A company deciding whether AI can handle 50,000 internal tasks cares about whether the model works, but accounting cares whether it works cheaply enough.
Anthropic also reports stronger behavior in its own tests: Opus 5.5 attempted to cross containment boundaries about 85% less often than Opus 5 or Mythos 5.1, with all remaining attempts rated low severity and self-reported. External evaluators including METR and Frontier Design tested the model before release, though company evaluations are not universal proof. The commercially useful question is becoming: how cheaply can we make it smart enough?
Source: Anthropic ↗Apple · Mac · Local AI
Apple's answer to token pricing is apparently “buy the computer once”
Apple's upgraded Mac mini and Mac Studio systems began shipping Tuesday with an unusual enterprise pitch: some intensive AI work may be cheaper on owned hardware than on rented cloud capacity. High-end configurations can approach $20,000, but they avoid a charge for every token and can keep sensitive work local.
At its launch event, Apple demonstrated four networked Mac Studios running a trillion-parameter model that found and repaired a graphics-programming bug. The four-machine cluster operated from one wall outlet. Unified memory and custom networking help the machines handle large local models, turning a design rooted in power efficiency into an advantage for AI inference.
Apple still holds only about 4.6% of the enterprise computer market, compared with 91.3% for Windows, according to IDC figures cited by Reuters. Microsoft is pursuing the same broad idea of on-device “unmetered intelligence.” Even so, capable local hardware pushes against years of AI centralization: keep the data nearby, avoid network delay and pay for the machine once. The cloud is discovering competition from the electrical outlet.
Source: Reuters ↗Geely · EVs · Fast charging
The charging stop is approaching “barely enough time to buy coffee”
Geely unveiled what it calls the industry's fastest EV charging system Wednesday. Supported by 2.2-megawatt infrastructure and a high-rate battery, the company says it can charge a vehicle from 10% to 70% in 4.5 minutes and reach 97% in 8 minutes 40 seconds under normal ambient temperatures.
Those figures approach a psychological threshold. One persistent objection to electric vehicles is waiting: gasoline adds hundreds of kilometers of range in minutes, while early EVs often needed half an hour or longer. Single-digit charging times make the experience feel much more familiar, although Geely's Zhang Dewang says the industry is approaching physical limits and future speed gains may diminish.
The obstacle is infrastructure. Delivering 2.2 megawatts at one connection demands an electrical system closer to industrial equipment than a household charger. Geely says cloud-based digital twins and deep-learning models inside vehicles monitor battery health, control temperature and predict hazards. At 4.5 minutes, the biggest charging problem may be returning before the car is finished.
Source: Reuters ↗