Apple Explores Third-Party AI Models for Siri Amid Internal Struggles

Apple is exploring the integration of third-party language models (LLMs) into Siri, as part of a strategic shift in its AI strategy. The company is in advanced discussions with Anthropic and OpenAI, following earlier talks with Perplexity and Thinking Machines Lab. This move comes amid growing concerns that Apple's in-house AI models, known as the "Apple Foundation Models," are not meeting expectations, particularly when compared to the rapidly advancing capabilities of generative AI products like Anthropic’s Claude and OpenAI’s ChatGPT.

Internal testing has shown that Anthropic’s Claude currently outperforms Apple’s own models. The evaluation of third-party models began after a management disruption shifted Siri responsibilities from the AI chief, John Giannandrea, to the software engineering head, Craig Federighi, and the Vision Pro executive, Mike Rockwell. Rockwell, who assumed the Siri engineering role in March, initiated the testing of external LLMs, including Claude, ChatGPT, and Google’s Gemini, to assess their performance against Apple’s in-house systems.

After several testing phases, executives found Anthropic’s Claude to be the most promising candidate for Siri’s requirements. This led Adrian Perica, the vice president of corporate development at Apple, to begin negotiations with Anthropic. Discussions have included requests for custom versions of Claude and ChatGPT to be trained and deployed on Apple’s Private Cloud Compute servers. However, Anthropic is reportedly seeking a multibillion-dollar annual licensing fee that would increase significantly over time, complicating the talks.

The potential switch to third-party models is notable because Apple has long prided itself on end-to-end control over its products, especially in matters of privacy and user experience. Siri already uses ChatGPT to handle some search-based queries and upcoming image generation features in iOS 26, but the assistant itself remains built on Apple’s models. OpenAI had previously offered to develop on-device models for Apple, but the company declined at the time. Until now, the company was expected to launch an upgraded Siri in 2026, powered by its internal models. These plans remain in progress under a parallel initiative known internally as “LLM Siri.”

While Apple’s discussions with potential partners seem to be heading in the right direction, internal issues appear to be demoralizing workers. A group of engineers led by Ruoming Pang and reporting to senior AI director Daphne Luong, who work on Apple’s Foundation Models, have reported feeling demoralized by the company’s outside partnerships. They feel that exploring third-party LLMs places blame on them for Apple’s AI shortcomings. Some have even considered leaving for more lucrative offers from rivals, who are reportedly offering annual compensation packages ranging from $10M to over $40M.

Employees at Apple also have the rising concern that the company’s reliance on third-party models for Siri may eventually extend to other AI features, which could potentially undermine any long-term investments in Apple’s in-house technology. Tom Gunter, a senior researcher involved in Apple’s large language model development, left the company last week after eight years. The MLX team that developed Apple’s core open-source system for machine learning on its chips also threatened to resign. After internal negotiations and counteroffers however, they agreed to stay.

Apple has shelved some internal projects entirely. For instance, Swift Assist, an Apple-developed LLM tool for code generation in Xcode, was canceled about a month ago. Instead, Apple plans to integrate third-party coding models, including ChatGPT and Claude, into the new version of Xcode expected later this year. Apple executives are also reportedly divided about how far the company should go in relying on outside AI vendors. Rockwell and Federighi appear increasingly open to short-term third-party integration, but they also believe Apple should retain ownership of AI models in the long run, given their central role in future products like robotics and wearable devices.