Apple's new CEO is waging an AI campaign on two fronts simultaneously, and the logic connecting them matters more than either front alone. John Ternus held a closed-door meeting with software chief Craig Federighi in recent weeks, pressing him to accelerate Apple's roadmap and move beyond incremental Siri improvements toward genuinely agentic capabilities1,4,5,6. Days later, Apple's research team published a paper on arXiv detailing HALO, an autonomous humanoid robot control system that achieved a 91% average success rate across 360 single-task tests in six unfamiliar real-world environments3. The two moves share a common architectural aspiration: that Apple can eventually own the full stack from language-model reasoning to physical actuation, routing intelligence through its own silicon and software layers. HALO does not prove that thesis—it runs Alibaba's Qwen3.6-35B-A3B on a Mac mini controlling Unitree's G1 body, so Apple built neither the model nor the robot. What it does prove is that Apple's research org is actively prototyping the systems-integration layer that would connect reasoning to actuation, and that Apple silicon can already serve as the inference substrate. The gap between prototyping integration and owning every layer is the central risk in Ternus's bet.
The thesis is not that Apple is doing two things at once. Every large platform company is. The thesis is that Ternus is staking Apple's AI identity on a vertical-integration play—from cloud-to-device software agents down to embodied robotics—and that the HALO paper reveals how far the research org has moved toward the systems-integration layer of that goal, even as the foundation-model layer remains outsourced and the Siri team is still catching up. No source links the Ternus-Federighi meeting to HALO directly; the connection is architectural and strategic, not organizational. We state it as inference, not reporting.
Start with the Siri pressure. Bloomberg's Mark Gurman, citing anonymous sources, reported that Ternus is frustrated by what he sees as a reluctance to embrace AI with the same urgency as competitors. Federighi, described by Gurman as having concerns about AI's privacy, security, societal, and employment implications, has nonetheless delivered a functional version of Siri AI on what Gurman called an incredibly short timeline. But the company reportedly has no illusions that it is keeping pace with rivals. Two concrete improvements are expected in the coming months: fine-grained control over first-party apps (rotating and cropping images in Photos, organizing Safari tabs) targeted for iOS 27.3, and integration with third-party apps, letting Siri place orders through services like Grubhub or book reservations through OpenTable, expected to land in the spring. Gurman wrote that Apple could make Siri function more like an agent as early as next year, tracking goals and pursuing objectives without the user initiating every step.
The competitive frame is explicit. Gurman named OpenAI's Dots, Meta's Muse, and Grok as the agentic benchmarks Apple is chasing7. Those products can keep track of goals, determine what needs to happen next, and continue pursuing objectives autonomously. 9to5Mac noted that Apple has built up frameworks over the years, like App Intents, and that with Apple a user hypothetically would not need to depend on as much computer-use ability, making certain tasks less expensive. That on-device integration route could reduce dependence on expensive computer-use emulation, but the agentic Siri roadmap remains early: the first-party app controls are targeted for iOS 27.3, and third-party integration is expected in the spring.
Now layer in HALO. Apple's research team submitted a paper titled "Perceptive Humanoid Motion from Simulation" to arXiv on November 4, unveiling a three-tier autonomous control system validated on Unitree Robotics' G1 humanoid platform. The top reasoning layer runs Alibaba's Qwen3.6-35B-A3B locally on a Mac mini. Across 360 single-task trials the system averaged 91% success, though composite tasks (three tested) dropped to 70%. The team simultaneously released code for data generation, model training, and deployment.
ANALYSIS The architectural choice is revealing. Apple did not build its own foundation model for HALO's reasoning layer; it used Alibaba's Qwen3.6-35B-A3B, a mixture-of-experts model priced on Alibaba's own cloud at $0.375 per million input tokens and $2.25 per million output tokens as of November 8. That model scores 20.35 on FrontierMath tiers 1–3 v2 and 27.08 on ArXivMath 2026_06 in independently verified benchmarks. By running it on a Mac mini rather than in the cloud, Apple demonstrated that its silicon can serve as the inference substrate for embodied AI, not just phone-based assistants. But the analogy to agentic Siri has limits. Real-time robotic control is a latency-bound, single-user workload; agentic Siri at scale must serve hundreds of millions of users running complex multi-step reasoning chains with persistent memory. The former is an edge-inference problem; the latter is a fleet-serving problem that still requires cloud capacity. What the HALO demo does establish is that Apple's silicon can handle the per-device inference slice—the on-device portion of a hybrid architecture where the phone or robot runs local planning while the cloud handles heavier reasoning and long-horizon memory.
Apple's foundation-models leader, Ruoming Pang, departed the company in September. Alibaba's post-training chief, Bowen Yu, and technical leader Lin Junyang both departed Alibaba on November 6. Gurman also reported that Mike Rockwell, vice president of Siri engineering, is still considering leaving his role or Apple altogether next year. HALO's dependence on Qwen sits alongside these departures uncomfortably. The pattern—ambitious systems-level research published while senior model and Siri talent departs or wavers—exposes the central vulnerability in the vertical-integration thesis. Apple's research bench is producing ambitious work, but the company's ability to power it with in-house models remains undemonstrated.
The strongest counterargument is that these two efforts are less connected than they appear. HALO is a research paper, not a product announcement. The 70% composite-task success rate suggests the system is nowhere near production reliability. The research team noted that new physical tasks still require manually written prompts and designed motion keyframes, and that the system relies mainly on visual perception and struggles to directly sense contact states. And the agentic Siri roadmap, even on Gurman's reading, puts the first concrete capabilities—first-party app control—at iOS 27.3, with third-party integration expected in the spring. A skeptic would say Ternus is pressuring Federighi precisely because Apple is behind, and HALO is an unrelated research exercise that market analysts in China are using to pump humanoid-robot concept stocks.
That reading is partly right, and we should be precise about what the connection is and is not. HALO is not a product. The connection to agentic Siri is therefore not architectural novelty but organizational intent: Apple is simultaneously pushing an agentic-assistant roadmap for its device base and publishing embodied-AI research that runs on its own silicon. The difference in output modality—app actions versus joint torques—is real, but both require the same core capability: on-device inference powering a reasoning-to-action pipeline. That is the vertical-integration thesis Ternus appears to be pushing. It remains an aspiration, not an achievement: Apple built neither the model nor the body in the HALO demo, and the evidence that it will build both is organizational momentum, not shipped product.
We previously reported that Apple's AI costs are hitting the income statement and that Ternus offered no doctrine to justify them. The Siri-plus-HALO framing begins to sketch that doctrine: Apple is not just spending to keep up with chatbot rivals but investing in the systems-integration layer of a full-stack agency capability that spans software and physical AI. The foundation-model layer—the part that would make the stack truly Apple-owned—remains the missing piece. Whether the spending produces returns depends on whether Apple can close that model gap before competitors like OpenAI (which closed an $11 billion round on November 6) and Meta (which committed $4 billion on October 27 and $95 billion the day before) outspend it on the very capability it lacks.
What to watch. First, whether iOS 27.3, expected in the coming months, actually ships the first-party app control capabilities Gurman described; that is the nearest testable milestone for agentic Siri. Second, whether Apple follows the HALO paper with any hardware partnership or acquisition in the robotics space, particularly with Unitree, which IPO'd in August. Third, whether Rockwell departs the Siri engineering VP role; Gurman reported he is still considering it for next year. A departure would signal deeper organizational friction beneath the Ternus-Federighi dynamic. Fourth, whether Apple develops or acquires its own foundation model for embodied AI rather than relying on Alibaba's Qwen line—a dependency that sits awkwardly with Apple's privacy and supply-chain positioning and that grows more precarious as Alibaba's own model team fractures (Bowen Yu and Lin Junyang both departed November 6).