AI Sector Daily Digest: July 25, 2026

In brief

1. Anthropic launches Opus 5 as a lower-cost frontier model

• Anthropic launched Claude Opus 5 for office work and programming, saying it approaches the capability of its more powerful Fable 5 model at half the price. 1
• The company says Opus 5 is more efficient than May's Opus 4.8 and showed lower cyber-exploitation capability than Fable 5 in testing, so its safeguards are less restrictive.
• Anthropic positions Opus 5 for value-focused, everyday work and Fable 5 for days-long autonomous projects. Reuters did not give an absolute dollar price for either model. 1

2. South Korea announces $950 billion AI and chip initiatives

• South Korea announced $950 billion in new AI initiatives involving Samsung Electronics, SK Group, and U.S. technology companies as chip demand continues to outrun supply. 2
• SK Group said its deals total $750 billion, including an SK Hynix partnership with Nvidia valued at more than $500 billion; Samsung separately signed a memorandum of understanding with Broadcom covering up to $200 billion across memory, foundry, and advanced packaging.
• The Nvidia and SK initiative includes a planned 2-gigawatt SK Telecom data center using Vera Rubin chips and SK Hynix HBM4 memory, targeted for 2027. These are announced deals and plans, not completed spending. 2

3. Anduril reportedly discusses a $100 billion valuation

• Anduril is in talks with investors on a new funding round that could value the defense-tech company at roughly $100 billion, Reuters reported, citing two people familiar with the matter. 3
• The amount to be raised is unknown, the valuation is still fluid, and one proposed structure would split the financing into two stages with a higher valuation tied to future financial benchmarks.
• Anduril said no decision has been made about future financing. The company last raised $5 billion at a $61 billion valuation two months ago, led by Thrive Capital and Andreessen Horowitz. 3

4. Indian court backs OpenAI in ANI copyright case

• An Indian court said OpenAI's use of news agency ANI content to train the models behind ChatGPT did not amount to copyright infringement. 4
• Justice Amit Bansal of the Delhi High Court said ANI had not shown that ChatGPT memorized or reproduced its reports in user responses, and treated the model-training storage as protected fair dealing for research under India's Copyright Act.
• The decision is an India-specific finding in a wider fight over training on copyrighted news. Similar cases against OpenAI in the United States and Canada remain separate proceedings. 4

5. Snorkel backs open benchmarks for long-horizon agents

• Snorkel AI announced the first projects supported by its Open Benchmarks Grants program, a $3 million commitment for open-source datasets, benchmarks, and evaluation research. The report carries a company-distributed release via HPCwire. 5
• The supported work includes Frontier-Bench for harder terminal tasks, Agents' Last Exam for professional workflows across 55 sub-industries, OSWorld 2.0 for computer-use agents, and benchmarks for continual learning and coding-agent quality.
• The program says applications remain open on a rolling basis. The announcement describes the projects and targets, but it does not provide independent results showing that any benchmark is already more predictive or reliable than existing evaluations. 5

What to watch