Anthropic's New Model Landed the Same Week It Bet $35 Billion on More Computing Power

Anthropic’s newest Claude showed up this week with an unusual companion: a receipt for computing power the company insists it will need for years to come.

Anthropic’s new Claude comes with a $35 billion insurance policy

On September 1, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, the company’s most capable models yet for coding, science and other knowledge work. The jump on some benchmarks is sharp. Fable 5.1 more than doubled Fable 5’s score on Terminal-Bench-Science, a test of scientific reasoning, climbing from 24.7 percent to 52.6 percent. Anthropic also cut the price of reading from its prompt cache by 75 percent, which it says makes heavily automated, agent-driven workloads up to 45 percent cheaper to run than before.

Mythos 5.1, the more powerful of the two models, is not being handed out freely. Anthropic is routing it through two restricted access programs built for cybersecurity researchers and biologists, the kind of specialists who need its sharper edges and are least likely to misuse them.

The launch landed a day after Anthropic signed a $35 billion computing deal with Lambda, a cloud provider backed by Nvidia. Under the arrangement, infrastructure company Hut 8 will build a roughly 350 megawatt data center in Nueces County, Texas, and Lambda will fill it with Nvidia chips that Anthropic will rent to run Claude. It is the fourth such megadeal Anthropic has struck in recent months, after commitments of 45 billion dollars with the cloud provider Nscale, 50 billion with Fluidstack, and 45 billion with SpaceX. Add them up and Anthropic has now committed more than 175 billion dollars to future computing capacity, a bet that demand for its models will keep outrunning the hardware built to serve them.

Alibaba is giving away what American labs charge a premium for

Alibaba took the opposite approach to the same problem. Its Qwen team released the weights for Qwen3.8-Flash-Next, which the company describes as an early preview of the architecture behind its next major model, Qwen4, free for anyone to download and run.

The model has 125 billion parameters in total, but only about 6 billion of them do any work on a given word, which is what keeps it fast and cheap to run despite its size. It has a native context window of 262,000 tokens, stretchable to a million, long enough to hold several full length novels at once. Developer access to Alibaba’s hosted version opened on September 1, priced at a fraction of what comparable American models charge. Where Anthropic is spending tens of billions to keep its best models fast and well resourced, Alibaba is betting that giving its work away builds a bigger following for whatever it charges for next.

A third of companies would rather build the software than buy it

A new McKinsey survey of 1,719 business leaders worldwide, part of its State of AI in 2026 report, found that 32 percent of organizations have decided against buying at least one piece of software because they judged they could build it themselves with AI coding tools instead. The share of large companies running AI agents across more than one business function jumped from 27 percent to 40 percent over the past year.

The finding says less about any one company than about how ordinary AI coding tools have become inside large organizations. Software vendors used to compete mainly with each other. Increasingly they are competing with a customer’s own engineering team and a chatbot.

Bihar just switched on enough power for a small country

Away from the AI industry, state-run SJVN Limited completed a 72 hour trial run on the second of two 660 megawatt generating units at the Buxar Thermal Power Project in Chausa, Bihar, on August 31. The two units together make up a 1,320 megawatt plant built with supercritical technology, which burns coal more efficiently and with a smaller emissions footprint than the older plants of the same size that still make up most of India’s coal fleet. Once both units are fully commissioned, the project is expected to generate close to 9.8 billion units of electricity a year, enough to meaningfully ease the power shortages that have long held back industry in one of India’s poorer states.

The most prolific wicket taker in T20 history plays a format most fans still ignore

India’s Deepti Sharma became the leading wicket taker in the history of women’s T20 international cricket on August 31, reaching 171 wickets during a 94 run win over Thailand at the Women’s Asia Cup in Dubai. She got there in her 150th appearance for India, only the third Indian woman to reach that many matches in the format, after Harmanpreet Kaur and Smriti Mandhana. Her spell against Thailand ran just over one over, conceded nothing, and took three wickets, the kind of unspectacular, efficient bowling that has built the record one tournament at a time rather than in a single headline moment.

Off spin bowlers rarely draw the attention fast bowlers and big hitting batters get, and women’s T20 cricket still receives a fraction of the coverage the men’s game does. Sharma has spent a decade quietly building a case that neither of those facts should matter as much as they do.

Put side by side, this week’s AI stories describe an industry pulling in two directions at once. One lab is spending whatever it takes to keep its best models fast and exclusive. Another is giving its work away to win the crowd instead. Bihar and Deepti Sharma are working on slower, more patient timelines, adding the megawatts and the wickets one project and one match at a time. Neither kind of progress makes headlines by itself. This week, side by side, they did.