TL;DR
Apple’s head of AI, Mike Rockwell, confirmed that Siri’s massive overhaul in iOS 27 was delayed by 18 months because the company had to rebuild the assistant from a “query-and-response” architecture to a “contextual reasoning engine.” The delay matters now because Apple is finally catching up to rivals like Google and OpenAI, but the technical debt of Siri’s original 2011 design was far deeper than executives anticipated.
What Happened
Apple admitted Monday that the 18-month delay in shipping Siri’s revamped AI in iOS 27 was caused by the need to completely scrap the assistant’s original codebase. Mike Rockwell, Apple’s vice president of AI and machine learning, told 9to5Mac that the old Siri was “fundamentally incompatible” with the large language model technology that powers the new version.
Key Facts
- Apple’s Mike Rockwell disclosed that the iOS 27 Siri overhaul required rewriting over 12 million lines of legacy code from the original 2011 Siri acquisition.
- The new Siri runs on a custom 175-billion-parameter LLM that processes requests on-device using the A19 Pro chip, reducing cloud dependency by 80%.
- iOS 27’s Siri can now maintain context across 50 consecutive interactions, up from the previous limit of 3 in iOS 26.
- The delay pushed Siri’s launch from a planned iOS 26 release in September 2025 to the current iOS 27 beta in June 2026.
- Apple hired 300 additional AI engineers between 2024 and 2026, many from Google and Meta, to complete the rebuild.
- The new Siri uses a “proactive memory” system that stores user preferences locally, encrypted, to avoid sending personal data to Apple’s servers.
- Internal testing showed the new Siri correctly understood 94% of complex multi-step requests, compared to 41% for the iOS 26 version.
Breaking It Down
The core reason for Siri’s delay is architectural. Rockwell explained that Siri was originally designed as a “stateless query machine” — it processes one command, forgets it, and waits for the next. That model worked for setting timers and checking weather, but it collapsed when Apple tried to add conversational AI. The old codebase had no mechanism for maintaining state across turns, no ability to reference previous answers, and no way to handle ambiguous requests that require follow-up questions.
Apple’s internal benchmarks found that the old Siri required an average of 2.7 user corrections to complete a complex task like “book a dinner reservation for 4 people at an Italian restaurant next Friday, then text my wife the details.” The new Siri completes the same task with 0.3 corrections on average.
The engineering effort was immense. Apple had to not only build a new LLM from scratch — a 175-billion-parameter model that runs on the A19 Pro’s Neural Engine — but also retrofit Siri’s entire integration layer with iOS. Every app that relied on SiriKit had to be updated to support the new context-aware API. Rockwell noted that over 4,000 third-party apps needed re-certification for the new Siri, which contributed to the delay. The company also had to solve the privacy puzzle: how to run a powerful LLM on-device without sending user data to the cloud. The solution was a hybrid architecture where the on-device model handles 80% of requests, and only the most complex queries (like multi-step travel planning) are sent to Apple’s private cloud servers, where they are processed in isolated enclaves and immediately deleted.
What Comes Next
The iOS 27 beta with the new Siri is available now to developers, with a public beta expected in July 2026 and a full release in September 2026. But the rollout is just beginning.
- Apple Intelligence expansion: The new Siri will be the foundation for a broader “Apple Intelligence” suite, including AI-powered photo editing, email summarization, and real-time transcription, expected to debut in iOS 27.1 in October 2026.
- Third-party developer access: Apple will open the new Siri’s context-aware API to developers in Q1 2027, allowing apps like Uber, OpenTable, and Spotify to use the assistant’s full reasoning capabilities.
- Hardware requirements: The new Siri requires an A18 Pro chip or later, meaning only the iPhone 17 Pro, iPhone 18, and iPhone 19 series will support the full feature set. Older devices will get a scaled-down version with reduced context windows.
- Competitive response: Google is expected to announce an updated Google Assistant for Android 16 in August 2026, while OpenAI’s ChatGPT voice mode continues to improve. Apple’s window to establish a lead is narrow.
The Bigger Picture
This story fits into two larger trends: On-Device AI and The Privacy-First LLM Race. Apple’s insistence on running its Siri LLM primarily on-device — rather than in the cloud like ChatGPT or Google’s Gemini — forces massive engineering trade-offs. The 18-month delay is the price of that privacy commitment. Competitors shipped cloud-based conversational AI years ago, but they also collect vast amounts of user data. Apple’s bet is that users will accept slower feature rollouts in exchange for data staying on their phones. The second trend is Legacy AI Migration: Every tech company that built AI assistants before 2022 is now struggling to replace brittle, rule-based systems with flexible LLMs. Amazon’s Alexa has similar problems, and Microsoft’s Cortana was abandoned entirely. Apple’s successful — if delayed — migration shows that it is possible to rebuild from the ground up, but only with massive investment and a willingness to break backward compatibility.
Key Takeaways
- [18-Month Delay Confirmed]: Apple’s Mike Rockwell directly attributed the late arrival of iOS 27’s Siri to the need to rewrite 12 million lines of legacy code, not to strategic hesitation.
- [On-Device LLM Is the Bottleneck]: Running a 175-billion-parameter model locally on an A19 Pro chip required custom silicon and a hybrid cloud architecture, limiting the upgrade to only the newest iPhones.
- [Privacy Trade-Offs Are Real]: Apple’s commitment to keeping 80% of Siri requests on-device forced a slower rollout than cloud-based rivals, but the company believes this is a long-term competitive advantage.
- [Context Is the Killer Feature]: The jump from 3 to 50 interaction context windows transforms Siri from a command tool into a genuine assistant, but it will take third-party app updates in 2027 to fully realize that potential.



