Around this time each year, when a new flagship chip debuts, the first question is almost always: what's the benchmark score? We spent time with a Dimensity 9600 Pro engineering unit, and the figure that truly stood out wasn't the score at all—it was power consumption.

According to MediaTek, this is the industry's first smartphone chip built on an "AI-native architecture." This isn't just adding more NPU muscle; the entire SoC—from CPU, GPU, and ISP to NPU—underwent a full modular redesign around AI at the blueprint stage. The result is an efficiency boost across the entire AI pipeline, smoother coordination between computing units, faster on-device large model inference, and improved performance-per-watt, all while delivering more stable output and lower latency in complex multimodal AI scenarios. On the experience side, it shifts the phone from a reactive "you call, it acts" model to one that proactively senses and handles tasks for you.

For this piece, benchmarks and gaming were hands-on; AI and imaging features couldn't be tested on the engineering unit, so we'll explain the official logic behind those and share our take at the end.

The silicon foundation: 2nm has arrived

Let's lay out the basics first. The Dimensity 9600 Pro is built on TSMC's 2nm process, packing 33 billion transistors. MediaTek's tagline this generation is "2nm means Dimensity," signaling not just early access to TSMC's leading-edge node but deep co-engineering in design and manufacturing to squeeze out maximum performance and efficiency. That said, the process node is just the foundation; what truly defines the experience is how MediaTek leveraged 2nm in the chip's design.

CPU: Rebuilt for multitasking and agents

Starting with the CPU, the short version is that the Dimensity 9600 Pro essentially rebuilds the CPU for the agentic smartphone era. The previous 9500 used a 1+3+4 setup; this generation shifts to 2+3+3, adding a second super-large core. The reasoning is straightforward: the era of AI agents is here. Some might wonder—doesn't the NPU handle large models? Why stack CPU super cores? The key lies in the agent itself: the NPU generates outputs, but background task orchestration and tool-calling logic fall to the CPU. When an agent runs in the background while a heavy app runs in the foreground, multiple high-priority threads compete, and a single super core struggles to keep up. The accompanying cache, memory, and storage specs have also been maxed out.

In Geekbench 6, CPU performance is up around 15%, which is standard for a flagship generational leap. The real standout is power: official figures show single-core power draw down 37% and multi-core down a massive 61%. How does that work with an extra super core? Previously, one super core had to brute-force its way through multiple heavy workloads, pushing frequencies to extremes and causing power to spike. Now, two cores share the load in their most efficient frequency ranges, avoiding the need to max out every time. Combined with the 2nm process, larger caches, and a new scheduling engine, performance climbs steadily while power drops sharply—that's why I call it a "rebuild."

GPU: Neural rendering and gaming experience

The CPU was a rebuild, and the GPU this generation brought equally impressive surprises, especially for gamers. The new architecture completely revamps the rendering pipeline. In 3DMark, scores are up 26%, with ray tracing up 19%, matching official claims. Ray tracing deserves extra attention. The bulk of ray tracing compute involves finding where light lands in a scene, and the third-generation engine speeds up this "ray-hunting" step by 18%. More critically, part of the previous generation's ray tracing relied on software; this time it's fully baked into the hardware. To put it simply, last year was like translating with a dictionary in hand; this year, the dictionary is in your head—ray tracing can finally run within a phone's power budget. The GPU also gains neural network fusion rendering, supporting game super-resolution that upscales from 540p to 1080p, with AI-optimized image quality at minimal extra power draw. The fourth-generation frame interpolation supports up to esports-level 185 fps while saving around 20% power.

Now, the hands-on gaming tests, with power as the focus. First up, Honkai: Star Rail at 60 fps with the highest graphics: at full frame rate, power draw was about 10% lower than the previous engineering unit. This isn't saving power by dropping frames—it's running at the same high frame rate while drawing less power, meaning you get the same smoothness, a cooler device, and better battery life. Second, Genshin Impact at max settings: nearly a flat line at full frame rate, with total device power at 4.0W—impressive by any measure.

One observation from my side: some in the industry are aggressively stacking GPU cores, with theoretical power draw at full load reaching 20W. But phones lack active cooling; the real sustained power ceiling is around 5-6W. A 20W spec is meaningless inside a phone. MediaTek chose a different path this generation—not chasing unattainable peak numbers with core counts, but focusing on performance-per-watt within the power wall. For real-world gaming, that approach actually works. The Dimensity 9600 Pro isn't chasing paper specifications; it's delivering performance users can actually use, turning new CPU and GPU hardware and software into a top-tier gaming experience. That's what justifies the "Pro" label.

NPU: Dual NPU architecture for proactive agents

Moving to the NPU, as the AI-native execution core, its job is to serve "proactive agents." The old model was "you call, it acts"; the new model keeps the system sensing in the background, proactively pushing services to you. This requires two capabilities: fast heavy lifting and always-on operation without draining the battery. The Dimensity 9600 Pro therefore uses a dual NPU architecture: a high-performance NPU handles complex generation and heavy inference, with peak performance up 51%, while a high-efficiency NPU sits in a "low-power island," on standby 24/7 with power draw 40% lower than the previous generation. Tasks like voice-to-text transcription and meeting notes are offloaded to this NPU, letting it run all day without the phone overheating. The most tangible figure: Tokens generated per watt of power is up 55%. In AI, it always circles back to efficiency—an agent that's always online must not overheat, so power efficiency decides the winner.

Beyond efficiency gains, peak performance is also top-tier: the Dimensity 9600 Pro tops the Zurich AI performance leaderboard, achieving breakthroughs in both AI performance and power efficiency. To be transparent, the engineering unit can't run on-device generation tests, so we couldn't verify proactive agent behavior in a real system—that will have to wait for the production unit.

Imaging ISP: A big year for video with AI

For imaging, this generation marks MediaTek's "video year." In still photography, Android vendors are already in the top tier, so this ISP focuses its firepower squarely on video. Dynamic range reaches 17 stops, preserving highlights and shadow details in high-contrast scenes. AI and ISP are now deeply integrated: previously, smart scene recognition topped out at 4K30; this generation debuts CSISP, pushing cinematic-quality 4K60 video with separate tonal and color processing for faces, hair, and skies. Multi-camera switching maintains color consistency, and tracking autofocus doubles to 60 frames per second. The technical specs for video on the Dimensity 9600 Pro are impressive on paper; we'll verify the actual output with production units.

Display and peripherals: The power-saving philosophy in the details

Quick rundown of display and peripheral upgrades: low-light flicker-free display keeps small text and subtitles clear, and H.266 video format support includes hardware decoding—future-proof for high-spec streams where hardware decoding saves power over software decoding. It all loops back to efficiency, the theme running through the entire chip.

Bottom line: Composure is the differentiator

Having spent time with it, here's my take. The Dimensity 9600 Pro's most lasting impression is its comprehensive reduction in power draw across all scenarios. Over the years, we've seen core-count wars and benchmark inflation, and everyone has felt the heat, throttling, and frame drops that follow overclocking. The Dimensity 9600 Pro takes a user-experience-first route: given a phone's real power budget of 5-7W, it maximizes real-world experience within that limit. In the first year of 2nm, the true dividing line for flagship chips isn't who peaks harder—it's who stays cooler and more stable in daily use and gaming. The Dimensity 9600 Pro's AI-native architecture represents the direction of flagship chip design in the AI era, from the revamped CPU and AI-accelerated GPU rendering to the dual NPU setup and AI-driven ISP. All of it is oriented toward proactive agentic experiences—this is what a new-generation flagship should be.