When it comes to in-house innovation, few companies are as proud as Apple. Yet, behind the scenes, the iPhone maker has quietly confronted a reality check: its long-promised, generative AI–powered Siri simply wasn’t ready for prime time. According to Bloomberg’s Mark Gurman, Apple has begun testing third‑party large language models (LLMs) from OpenAI and Anthropic on its private cloud—an extraordinary reversal for a company that usually tightens control over every line of code.

The development of “LLM Siri” has been a rocky ride. Originally scheduled to make its debut earlier this year, the overhauled assistant was delayed in March amid quality‑control concerns. At Apple’s Worldwide Developers Conference (WWDC) in June, SVP of Worldwide Marketing Greg Joswiak admitted candidly that the system “didn’t hit our quality standard,” underscoring just how far Apple still has to go.

Behind the scenes, Apple’s AI and Siri division had been helmed by John Giannandrea, a veteran AI executive hired from Google in 2018. But as delays mounted and internal pressure rose, CEO Tim Cook reportedly “lost confidence” in Giannandrea’s leadership, prompting a shakeup. In his place, Cook tapped Mike Rockwell—best known for steering the Vision Pro project—to spearhead Siri’s renaissance.

Under Rockwell’s watchful eye, Apple invited OpenAI and Anthropic to tailor versions of their flagship models—ChatGPT and Claude, respectively—for evaluation on Apple’s servers. Google’s Gemini was included for comparison, though its use would likely be more restricted given Google’s status as a direct competitor in both hardware and AI services.

Early feedback has reportedly favored Anthropic’s Claude, which Apple engineers found to be the most promising at handling everyday requests and striking the right balance between brevity, accuracy, and privacy. OpenAI’s GPT‑4–powered ChatGPT also impressed, particularly on more conversational tasks. Meanwhile, Google’s Gemini, already baked into Android and the Pixel lineup, served as a useful benchmark—reminding Apple just how far behind it had fallen in consumer‑facing AI features.

Apple has long campaigned on user privacy as its north star. Running third‑party models on its private cloud—powered by Apple silicon and end‑to‑end encryption—allows it to offer the computational heft of GPT‑4 or Claude without compromising on-device confidentiality. In theory, sensitive voice data would never leave Apple’s servers unencrypted, preserving the privacy stance that differentiates Siri from rival assistants.

Across the tech landscape, alliances and licensing deals are reshaping the AI battleground. Samsung, for instance, licenses Google’s AI technologies to supercharge its Galaxy phones. Motorola is rumored to be close to sealing a similar pact with Perplexity, the AI search engine that some Apple executives had even considered acquiring earlier this month.

Google itself has been ramping up Gemini integrations on Android, while Microsoft continues to pour resources into its partnership with OpenAI—embedding ChatGPT features across Windows and Office. Meta, for its part, pushes forward with its open‑source Llama models. Even Amazon has thrown its hat in the ring with homegrown models like Nova and Titan for AWS customers. Amid this frenzy, Apple’s late‑arrival strategy of testing multiple partners could pay dividends—if it can move swiftly.

Apple’s exploration remains “in the early stages,” according to Bloomberg. No final decision has been announced, and Apple still holds out the option to double down on refining its own foundation models. Still, the willingness to consider outside technology marks a profound cultural shift—one that acknowledges Siri’s struggle to keep pace with AI leaders.