Apple Partners with Anthropic to Launch ‘Vibe-Coding’ Platform

Apple has partnered with Anthropic to create a “vibe-coding” platform for automating code generation, editing, and testing. The brand new system, designed to reinforce Apple’s Xcode, will integrate Anthropic’s Claude Sonnet AI model. “Vibe coding,” which leverages AI agents to jot down code, is gaining momentum within the industry. While Apple plans to make use of the platform internally, a public release is yet to be decided. This collaboration follows delays with Apple’s Swift Assist AI tool for Xcode, furthering its push to include AI into its devices.

In a separate development, Anthropic is offering current and former employees the chance to money out as much as 20% of their shares, with a cap of $2 million per person. This buyback, valued at $61.5 billion based on its March funding round, applies to staff who’ve been with the corporate for a minimum of two years. Anthropic, founded by former OpenAI researchers and now employing over 800 people, is anticipated to finish the buyback by month’s end.

Anthropic has upgraded Claude’s research mode, allowing it to run investigations for as much as 45 minutes—triple the prior limit—before producing document-style reports. The feature now dissects complex queries into smaller components and scours “tons of of internal and external sources,” with full citations embedded. Users can expect reports to surface nuanced details, though AI-generated content still requires sharp-eyed fact-checking; in testing, Claude’s historical report on video games was impressively thorough but included a confabulated quote.

Anthropic also rolled out its recent “Integrations” feature, linking Claude with tools like Jira, Zapier, and PayPal via the Model Context Protocol. The update broadens Claude’s data access beyond web search and Google Workspace, enabling automations similar to compiling HubSpot sales data or generating Jira tickets in bulk.

Each the upgraded research mode and Integrations are actually available to all paying users, with Pro accounts expected soon.

DeepMind’s Nikolay Savinov explains that despite the promise of million-token context windows, specializing in essentially the most relevant content yields higher results. In a recent interview, Savinov noted that when an AI model processes large amounts of knowledge, it must divide its attention across all tokens, potentially diminishing deal with essential details.

To optimize accuracy, Savinov suggests trimming documents all the way down to only the needed information. For instance, removing irrelevant pages from a PDF before sending it to an AI model can result in improved performance, even when the model is able to processing the complete document. Read more.