Skip to content

Weekly digest

25 – 31 May 2026

3 posts

Opus 4.8 on its low effort setting uses as many output tokens on SWE-Bench Pro as the medium or high settings did in Opus 4.7 and 4.6, according to system card data. Medium in 4.8 consumes more than high in 4.7 and approaches the absolute maximum of 4.6. Higher token consumption directly affects API speed and cost.

Apollo and Blackstone plan to raise $36 billion through a special entity to buy Google TPU chips and lease them to Anthropic. The arrangement lets Anthropic access vast compute without taking on massive debt directly.

What it takes to work at Anthropic

Anthropic has become one of the most coveted employers in the industry, with salaries often exceeding $250,000 per year. But landing a job there requires more than just exceptional intelligence or an elite resume.

Commercial Director Paul Smith looks for unconventional candidates who can fully immerse themselves in the client’s world. The ideal employee should help company leaders understand how AI fits into their workflows.

Beyond professional skills, Anthropic prioritizes cultural fit. The company seeks people who share its mission and can adapt to rapid growth.

Hacker NewsDiscuss

Apollo and Blackstone raise $36 billion for Anthropic’s chips

Investment giants Apollo and Blackstone plan to raise $36 billion to buy Google processors. The equipment is destined for startup Anthropic.

The deal will be done through a special entity. The firms will buy Google’s TPU chips and then lease them to Anthropic. This lets the startup access vast compute without taking on massive debt directly.

For Google, it means a large contract for in-house chip sales. Anthropic gets infrastructure to train next-gen models to compete with OpenAI.

Hacker NewsDiscuss

Effort levels recalibrated in Opus 4.8

SWE-Bench Pro system card data shows a major shift in token consumption. The low-effort setting in Opus 4.8 now uses as many output tokens as the medium or high settings in 4.7 and 4.6.

The efficiency of Opus 4.8 on low effort matches the high level of 4.7. The medium setting in 4.8 consumes more tokens than the high setting in 4.7, approaching the absolute maximum of 4.6.

Developers should reassess generation settings. Increased token consumption directly impacts API speed and cost.

Import: textDiscuss