AI Daily · 2026-07-25

Anthropic launched Claude Opus 5, delivering near-Fable performance at half the price with stronger prompt injection resistance, reshaping the cost‑capability landscape—community tests even show it surpassing Fable on certain agent benchmarks. The same day, SK Group and NVIDIA announced a planned strategic partnership exceeding $500 billion, spanning AI factories to memory supply, while South Korea’s national AI factory triples its planned capacity to 200 MW, underscoring a global infrastructure arms race. Separately, Anthropic’s Project Pilot demonstrated rapidly advancing autonomous drone control, highlighting both capability leaps and governance challenges.

North America · First-hand

Anthropic

⭐⭐⭐⭐ [Model Release] 2.1.220

Claude Code releases version 2.1.220, introducing Claude Opus 5 as the new default Opus model (1M context, fast mode) and removing Opus 4.7 from fast mode. New features include sandbox network strict allowlist, DirectoryAdded hook, and nested subagent depth increased to 3 by default, along with numerous bug fixes such as improved MCP connection error messages and screen-reader mode enhancements.

⭐⭐⭐ [Research] Project Pilot: Can AI control a drone?

Anthropic and Andon Labs present Project Pilot, testing frontier AI models' ability to autonomously control a quad-rotor drone for a locate-and-follow surveillance task, and introduce the Drone-Bench benchmark. The task requires the model to chain sub-tasks—flight control, obstacle-aware mapping, person identification from a reference photo, and real-time tracking— showcasing reasoning and execution improvements from prior projects like Project Fetch. The report emphasizes the dual-use nature of drone autonomy, the rapid capability gains, and the urgent need for governance norms among developers, civil society, and governments.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐⭐ [Model Release] [AINews] Claude Opus 5: Fable-level performance at Opus price (half Fable)

Anthropic launched Claude Opus 5, which officially "comes close" to Fable-level performance but independent benchmarks like AA-Briefcase show it already outperforming Fable with 20% lower cost per task. Epoch's ECI gives Opus 5 a score of 159 vs. Fable's 161, while tying on SWE-ECI. Community feedback highlights strong practical gains in coding and browser agent tasks, suggesting real-world improvements beyond what some benchmarks capture. The model delivers near-Fable capability at half the Fable price, offering a notable efficiency upgrade.

⭐⭐⭐ [Model Release] Quoting Boris Cherny

Boris Cherny of Anthropic highlights that Opus 5 is their most prompt-injection-resistant model yet. Red teaming and PI evaluation results in the system card show it is very difficult to inject successfully. This quotation was collected by Simon Willison on July 25, 2026.

⭐⭐⭐ [Model Release] Introducing Claude Opus 5

Simon Willison covered the release of Anthropic's Claude Opus 5, a model positioned as close to frontier intelligence at half the cost of Claude Fable 5. It currently leads the Artificial Analysis leaderboard, retains the same pricing and fast mode as Opus 4.8, and shows remarkable proactivity—such as writing its own computer vision pipeline to reconstruct a part from pixels when denied direct access. The model improved at finding cybersecurity vulnerabilities but was not trained on exploiting them. Anthropic also published a prompting guide and a context engineering article. Willison hasn't tested it thoroughly yet, sharing initial observations from the announcement.

Product Update

⭐⭐⭐⭐ [Product Update] SK Group and NVIDIA Expand Strategic Partnership Across AI Factories and Next-Generation Memory

SK Group and NVIDIA announced plans for a comprehensive partnership exceeding $500 billion to build AI infrastructure for surging global compute demand. The two companies signed letters of intent covering AI factory construction and next-generation AI memory supply. The collaboration aims to integrate NVIDIA's accelerated computing with SK Group's semiconductor and energy capabilities, marking an unprecedented investment scale in AI infrastructure.

⭐⭐ [Product Update] NAVER, NVIDIA and Brookfield to Expand Korea’s National AI Factory Infrastructure Buildout

NAVER, NVIDIA and Brookfield announced plans to expand Korea’s sovereign AI factory infrastructure from the initial 55-megawatt deployment to 200 megawatts, more than tripling the earlier buildout. NAVER further intends to scale its NVIDIA AI infrastructure deployment to 1 gigawatt. The collaboration aims to strengthen Korea’s sovereign AI capabilities through large-scale computing infrastructure.