重磅转向！传苹果拟弃自研 AI 模型，Siri 或求 OpenAI/Anthropic“外援” 迎头赶超

According to informed sources, Apple is considering adopting artificial intelligence technologies from Anthropic or OpenAI to drive a new version of Siri, planning to shelve its self-developed model in an effort to reverse its lag in the generative AI field. This decision could disrupt the industry landscape.

The iPhone manufacturer has been negotiating with both companies to upgrade Siri using large language models (LLM) and has requested them to train customized models compatible with Apple's cloud architecture for testing. If finalized, this would mark a significant strategic shift for Apple—currently, most of its AI features rely on its self-developed "Apple Foundation Model," which was originally planned to launch a new voice assistant based on this technology in 2026.

If Siri adopts Anthropic's Claude or OpenAI's ChatGPT model, it would undoubtedly be an acknowledgment of Apple's insufficient competitiveness in the generative AI field, a technology considered the most important innovation in decades. Although Apple has allowed ChatGPT to answer web search queries in Siri, the core of the voice assistant is still supported by self-developed technology. Informed sources indicate that the evaluation of third-party models is still in the early stages, and the self-developed project codenamed "LLM Siri" is still progressing.

The Ten-Year Dilemma of Siri and Management Reshuffle

The plan to evaluate third-party models was initiated by Siri head Mike Rockwell and software engineering chief Craig Federighi. Previously, AI head John Giannandrea was demoted due to the lackluster response to Apple Intelligence and delays in Siri upgrades, with the two taking over related responsibilities.

Rockwell, who led the development of the Vision Pro headset, took over the Siri engineering team in March this year and requested the team to assess the effectiveness of using Apple's self-developed models versus third-party technologies (including Claude, ChatGPT, and Google's Gemini). Multiple rounds of testing showed that Anthropic's technology best met Siri's needs, prompting Vice President of Corporate Development Adrian Perica to negotiate with Anthropic regarding collaboration on Claude.

Launched in 2011, Siri has fallen behind mainstream AI chatbots, and Apple's upgrade plans have repeatedly faced technical bottlenecks and delays. New features released last year (including accessing user personal data and screen content analysis) were originally planned to go live in early 2025 but have now been indefinitely postponed, with a current expectation for a spring 2026 launch.

AI Strategy Fluctuations and Talent Crisis

Apple's AI team is in a state of high uncertainty, with executives still weighing multiple paths. Although a multi-billion dollar budget for 2026 has been approved for running self-developed models in the cloud, long-term planning remains vague. Executives like Federighi and Rockwell are gradually agreeing that rather than relying on current self-developed technology, it may be better to leverage third-party solutions for short-term breakthroughs This strategy is similar to Samsung Electronics, although the functionality is branded as Galaxy AI, it is actually based on Google's Gemini. Anthropic has provided technical support for Amazon's new version of Alexa+. Executives believe that if self-developed technology improves in the future, given the importance of AI to product operations, Apple should still maintain ownership of the models. The company is currently advancing heavily AI projects such as desktop robots and smart glasses.

Apple's model is developed by a team of about 100 people, led by prominent engineer Ruoming Pang, who joined from Google in 2021, reporting to Senior Director of AI Research Daphne Luong. Luong is a core deputy of John Giannandrea, and the foundational model team is one of the few AI departments still reporting to Giannandrea, but Federighi and Rockwell have become more involved.

This shift has dampened team morale, with some members believing that the company's consideration of third-party technology undermines their work, and several have received multi-million dollar poaching offers from companies like Meta and OpenAI. Meta has offered annual salaries ranging from $10 million to $40 million for its newly established superintelligence lab, while Apple AI engineers typically earn only half of the market rate. Last week, Apple’s senior large language model researcher Tom Gunter left, and his unique skill set is hard to replace.

Technical Game and Business Discrepancies

In negotiations with Anthropic and OpenAI, Apple has requested customized versions of Claude and ChatGPT that can run on its private cloud computing servers—this architecture is based on high-end Mac chips and is currently used to run self-developed complex models. Apple believes that running models on its own servers can better protect user privacy and has completed internal feasibility testing.

Other Apple smart features use device-side models, which, although slower and less performant than cloud versions, can be used for tasks such as email summarization and generating custom emojis. Later this year, Apple will open device-side models to third-party developers, but has no plans to open cloud models yet, partly because server capacity currently cannot meet the demand for new features.

Insiders say that Apple and Anthropic have discrepancies over preliminary financial terms: the AI startup is asking for hundreds of millions of dollars annually, increasing year by year. The stalemate in negotiations has prompted Apple to consider that if third-party solutions are implemented, it may turn to collaborate with OpenAI or other companies.

If an agreement is reached, the influence of John Giannandrea, who joined from Google in 2018 and advocates for self-developed large language models, will be further weakened. In addition to losing control over Siri, his robotics department's responsibilities have also been stripped away, with the core machine learning and application intent teams now assigned to Federighi's software engineering department. The Swift Assist project, a code assistance tool announced by Apple last year, was also canceled a month ago, to be replaced by a new version of Xcode supporting ChatGPT/Claude, set to launch later this year