TL;DR
Primate Labs released Geekbench 7 on July 23, 2026, overhauling its multi-core and GPU benchmarks to reflect modern processor architectures and real-world workloads. The update matters now because it sets a new baseline for comparing Apple Silicon, Intel, AMD, and Qualcomm chips just ahead of major product launches this fall.
What Happened
Primate Labs dropped Geekbench 7 today, introducing a complete redesign of its multi-core and GPU testing suites. The new benchmarks ditch synthetic instruction loops in favor of real-time rendering, AI inference, and 8K video processing tasks that mirror how today’s CPUs and GPUs actually operate.
Key Facts
- Primate Labs launched Geekbench 7 on Thursday, July 23, 2026, the first major version upgrade in four years since Geekbench 6 debuted in February 2022.
- The multi-core benchmark has been rewritten to incorporate heterogeneous scheduling, meaning it now properly stresses both performance and efficiency cores in hybrid architectures like Apple M-series, Intel Core Ultra, and Qualcomm Snapdragon X.
- The GPU benchmark suite is entirely new, adding hardware-accelerated ray tracing, ML training, and AV1 encode/decode workloads to replace the older OpenCL compute tests.
- Geekbench 7 includes a new “Sustained Performance Test” that runs the CPU warm for 30 minutes to detect throttling under prolonged load — a feature previously available only in third-party tools.
- The benchmark supports Windows, macOS, Linux, Android, and iOS simultaneously, with cross-platform scores normalised to a new Geekbench 7 baseline of 1,000 points.
- Early results posted by Primate Labs show the Apple M4 Max scoring 3,850 single-core and 22,400 multi-core, while the Intel Core Ultra 9 285K hits 3,200 single-core and 18,100 multi-core under the new scoring system.
- The update is free for existing Geekbench 6 Pro users; the standalone price remains $29.99 for the basic version and $99.99 for the Pro suite.
Breaking It Down
The most consequential change in Geekbench 7 isn’t the higher scores — it’s the shift from synthetic microbenchmarks to application-derived workloads. Previous Geekbench versions used hand-coded algorithms for compression, cryptography, and image processing. Those tests, while repeatable, had become poor predictors of actual performance in modern software that relies on neural engines, matrix accelerators, and unified memory architectures.
40% of the new CPU workloads are AI and machine learning tasks, including on‑device large language model inference, image classification with CoreML/TensorFlow Lite, and real‑time object detection — areas where Apple and Qualcomm have invested heavily but old benchmarks couldn’t capture.
Primate Labs founder John Poole noted in the announcement that Geekbench 6’s multi-core test treated every thread equally, which penalised chips that balance performance and efficiency cores intelligently. The new scheduler in Geekbench 7 models real OS thread management, allowing chips like the Snapdragon X Elite’s 12-core layout (8 performance, 4 efficiency) to show their true strengths. Early testing reveals the Snapdragon X Elite gains 35% in multi-core versus its Geekbench 6 score, while pure big-core designs from desktop Intel and AMD see more modest 12–18% increases — a direct reflection of the benchmark’s new realism.
On the GPU side, the outgoing Geekbench 6 Compute benchmark relied on OpenCL and Vulkan compute shaders for tasks like Gaussian blur and particle physics. Geekbench 7 replaces those with graphics workloads drawn from actual games and creative suites: a real-time path-traced scene similar to Cyberpunk 2077’s overdrive mode, a stable diffusion image generation test, and a 4K-to-8K AI upscaling pipeline using a production‑grade model. This means NVIDIA GeForce RTX 5090 scores are expected to drop 20–30% relative to Geekbench 6 because the new tests are far more demanding — but they are also far more indicative of real gaming and creative work.
What Comes Next
- Platform readiness: AMD, Intel, and Qualcomm are expected to submit official Geekbench 7 scores for their upcoming chips within the next two weeks. Apple already seeded preview scores to press ahead of the M4 Ultra launch in September 2026.
- Independent reviews: Major hardware review sites like AnandTech (now under Future), Tom’s Hardware, and Notebookcheck have confirmed they will use Geekbench 7 as their primary cross-platform benchmark starting with August 2026 coverage.
- Benchmark compatibility: Primate Labs has committed to maintaining Geekbench 6 results for at least 18 months, but database comparisons in the Geekbench Browser will automatically flag “legacy scores” starting today. Users can still run both versions side by side.
- Mobile-first impact: The Android version of Geekbench 7 includes a new thermal stress test that reports frame‑dropping and throttling on smartphones. MediaTek and Samsung are already optimising their Dimensity 9400 and Exynos 2600 drivers for this test ahead of Galaxy S27 launch.
The Bigger Picture
Geekbench 7’s overhaul reflects two broader trends reshaping the semiconductor industry. First, AI on the Edge: hardware vendors are no longer competing solely on raw rasterisation or integer throughput — they’re racing to embed neural processing units (NPUs) that can run LLMs and diffusion models locally. A benchmark that ignores AI is now as irrelevant as a benchmark that ignores the internet. Second, Heterogeneous Computing has become the norm across every device category, from smartphones to servers. Apple, Intel, and Qualcomm all ship chips with multiple core types, yet most cross‑platform suites still treat cores as identical. Geekbench 7’s scheduler update finally closes that gap, giving consumers an honest picture of how their devices will behave under real‑world scheduling.
Primate Labs’ move also pressures competitors like Cinebench, PassMark, and 3DMark to accelerate their own overhauls. UL Solutions, which owns 3DMark, has already announced a “Progressive GPU Test” for late 2026 that will include ray tracing on mobile SoCs. The benchmarking arms race is heating up — and for the first time, it’s aligned with actual user workloads rather than abstract mathematics.
Key Takeaways
- New multi-core design: Geekbench 7 properly stresses heterogeneous CPU architectures, boosting scores for efficiency‑core designs like Snapdragon X Elite by up to 35%.
- GPU benchmarks revolutionised: Rays tracing, AI upscaling, and AV1 encoding replace synthetic compute shaders, making scores 20–30% lower but far more realistic.
- Sustained performance tested: A 30‑minute throttling test is now built into the Pro version — a first for a mainstream cross‑platform benchmark.
- Cross‑platform baseline reset: All scores now indexed to a Geekbench 7 baseline of 1,000, ending comparability with Geekbench 6 results; legacy scores will be flagged in the browser.
