Summary: A June 30, 2025 Bloomberg report said Apple was considering OpenAI or Anthropic models for the next Siri and had asked both labs to train versions for Apple’s own cloud infrastructure. The article frames the possible core-model change as a departure from Apple’s in-house Siri track, while emphasizing that the work was only an internal evaluation and no agreement had been announced.
On June 30, 2025, Bloomberg reported that Apple is considering using models from OpenAI or Anthropic to power the next version of Siri, and has asked both labs to train customized versions of their models that can run on Apple’s own cloud infrastructure for internal testing. TechCrunch and Reuters picked up the report the same day.
For a company that had staked its AI strategy on in-house models, this is a loosening at the level of strategy: the core model behind Siri may no longer come from Apple’s own labs. Timing matters here — the report broke the same day Meta’s superintelligence reorganization dominated the tech news cycle, a reminder of how wide the gap between frontier labs and even the wealthiest hardware companies had become by mid-2025.
Bloomberg’s report: custom models for Apple’s cloud
According to Bloomberg’s June 30 report — headlined around the idea of replacing Siri’s AI LLMs with Anthropic’s Claude or OpenAI’s ChatGPT — Apple has approached both labs separately, asking each to train model versions “that can run on Apple’s cloud infrastructure” for evaluation. Reuters carried the story the same day under the frame of a “major reversal.” The detail worth noting for technical readers: Apple is not looking to call a hosted API off the shelf, but to deploy customized models on infrastructure it controls, consistent with its long-standing constraints on privacy and data flow. Serving frontier-model quality inside that envelope is exactly the kind of engineering problem Apple has spent a year building toward — and exactly the kind the two vendors are now being asked to solve for Apple rather than for themselves. The effort remains at the evaluation and testing stage, and no agreement has been announced.
The split from the in-house “LLM Siri” track
Siri can already hand hard questions to ChatGPT, but that integration is shallow — a bolt-on escalation path rather than a new brain. The arrangement Bloomberg describes goes much further: third-party models would drive Siri itself, replacing the proprietary large language models that currently underpin the assistant. In other words, Apple’s earlier compromise outsourced the hardest questions while keeping the assistant itself on in-house models; the proposal on the table would move the core-model seat to an outside vendor. That marks a setback for Apple’s internal “LLM Siri” project, which bets on proprietary models, and it reflects the reality of a company that has spent years falling behind Google, OpenAI, and Anthropic in the AI race despite its distribution advantages.
The delay shadow and the clock
Apple originally slated its AI-powered Siri for 2025, then pushed it to 2026 or later after a series of reported technical challenges. The delayed, personalized Siri remains the biggest hole in Apple’s AI story, and the signal in the June 30 report is blunt: until its in-house models catch up, Apple is willing to seriously consider handing core capability to outside models and reclaiming time on its own terms. It also explains the form the evaluation takes — even while testing external models, Apple wants them running on infrastructure it controls, so data flows stay inside the company’s privacy rules. The ChatGPT hook in today’s Siri was always framed as a stopgap; the next-generation Siri was supposed to be the moment Apple’s own models took over. Whatever the outcome, the internal team now has a visible benchmark to race against.
What it means for model vendors and developers
For OpenAI and Anthropic, Siri may be the single largest consumer-AI distribution channel in history; the winner’s model would ship inside hundreds of millions of devices and shape hundreds of millions of users’ first-hand experience of an AI assistant. Even the evaluation stage signals the stakes: a deployment of this size would dwarf most enterprise contracts on distribution, if not on per-token economics. For developers, the most instructive part is the constraint: both labs were asked for versions that run on Apple’s own infrastructure. Beyond front-end competition, the ability to adapt model deployment to a customer’s privacy architecture is becoming a decisive factor for the biggest contracts. The talks are still at the testing stage — worth watching is whether Apple actually lands the deal, and whether “custom training for a specific customer’s infrastructure” becomes a product line of its own for model vendors.