Geekbench 7 vs 6: What Changed, Why Scores Differ

Geekbench 7: new tests, a new multi-core score and CUDA

Geekbench runs a fixed series of workloads, each modeled on a task from real software, and scores how quickly the device completes them. Version 7 brings new workloads, a rebuilt multi-core score, a GPU test refocused on machine learning and content creation, and larger data sets.

Geekbench 7 release date, platforms and download

Primate Labs announced the release on July 23, 2026, in a blog post by John Poole. Downloads opened the same day for Android, iOS, Windows, macOS and Linux. The post links to the download page for each platform, Google Play and App Store listings included.

Phones, tablets, Macs and PCs all get the same version, and a single app holds both the CPU and GPU benchmarks. The CPU run produces a single-core and a multi-core score, while a separate GPU run tests the graphics chip.

Pro costs $99 again after the launch discount ended

Personal use is free, and Primate Labs says it will remain so. The paid Geekbench 7 Pro is sold through the Primate Labs Store, and AppleInsider reports that it adds command-line tools and automated benchmark deployment for professional and commercial users.

Primate Labs launched Pro at 20% off until August 6, which AppleInsider put at $79 against the $99 list price. That offer is over: as of September 29, 2026, the Primate Labs Store lists Pro at $99, matching the price AppleInsider reported on July 23.

Geekbench 6 vs Geekbench 7 at a glance

Only the changes Primate Labs announced appear here, along with AppleInsider's verdict on scores. Because the announcement does not say what version 6 lacked, there is no Geekbench 6 column. The last row matters most if you still depend on Geekbench 6 scores.

New CPU workloads and the multi-core redesign

A single-threaded task uses one CPU core, whereas a multi-threaded task is divided so several cores work on it at the same time. Geekbench reports a single-core score for the first type and a multi-core score for the second. For the CPU, Geekbench 7 adds media and game tests, refreshes its photo tests and changes which workloads feed the multi-core score.

AV1, Opus and Whisper media workloads

Three new CPU workloads handle audio and video, modeled on video calls, screen sharing and streaming. Encoding compresses video or audio for sending or storing, and decoding unpacks it for playback. AV1 and Opus are codecs, or compression formats: AV1 for video, Opus for audio.

The AV1 test encodes the stream a conferencing app sends while you share your screen. The Opus test compresses music and speech recordings, as voice memo and podcast apps do. The third plays back video and audio while the Whisper speech recognition model generates live captions.

Of the three, only the captioning test runs an AI model. AppleInsider notes that it represents playback with automatic subtitles turned on, not speech recognition by itself.

Game Physics, Photo Editor and Photo Library updates

Outside media, Geekbench 7 adds one game test and reworks two photo tests. The game test uses the Jolt engine, which Primate Labs says appears in popular video games. Photo Editor now applies a broader set of real-world edits.

Photo Library now imports and processes modern image formats, with JPEG XL and DNG named in the post. Primate Labs presents both photo tests as expanded or updated versions of existing workloads rather than new ones.

Why multi-core no longer runs every test on every core

A workload is now spread across cores only when the software it imitates does so in everyday use. Primate Labs uses the HTML5 Browser test as its example. It is excluded from the multi-threaded suite because, according to the company, web browsers run on one thread or just a few.

Primate Labs argues that treating every task as parallel skews the score and tells you nothing useful about the device. It describes the result as a multi-core score that better reflects the work you actually do, and AppleInsider adds that the method behind the number now matters more when comparing devices.

The GPU benchmark shifts to machine learning and adds CUDA

An API is the programming interface an app uses to pass work to the GPU, such as Metal on Apple devices or CUDA on NVIDIA graphics cards. The GPU benchmark now centers on machine learning and content creation and adds CUDA as a fourth API. Primate Labs says these two kinds of apps increasingly decide GPU performance.

Machine learning workloads on the GPU

Three GPU workloads apply machine learning to video and images, each copying a feature people already use. One tracks faces in a video and adds live filter effects, as social media apps do. Another upscales images with machine learning, the super resolution feature in content creation apps.

The third blurs the background of a video call, like the virtual background option in meeting apps. The post does not describe these three as new, and Geekbench 6's GPU benchmark already included Background Blur and Face Detection tests.

Image editing and synthesis workloads

Four more GPU workloads deal with photos, video and simulated scenes. They are path tracing, fluid simulation, RAW photo processing and video color grading with LUTs. Path tracing draws a scene by following simulated rays of light, while fluid simulation calculates how a liquid or gas moves.

RAW processing converts a camera's unprocessed sensor data into a finished photo. A LUT, or lookup table, maps every input color to a new one so an editor can regrade footage in one pass. Photo, video and 3D apps all send these four jobs to the GPU.

CUDA joins OpenCL, Vulkan and Metal

Primate Labs says CUDA was added by popular demand, alongside OpenCL, Vulkan and Metal. CUDA is NVIDIA's GPU computing platform for its graphics cards. According to the post, owners of those cards can now test them through the API their most demanding apps use.

Nothing changes on Apple hardware. AppleInsider notes that Apple devices still use Metal.

Geekbench 6 scores don't carry over to Geekbench 7

A Geekbench score is a count of points, not a measurement in seconds or any physical unit, and AppleInsider notes it only means something within the version that produced it. Version 7 changed its workloads, data sets and multi-core method at once, so its scores cannot be compared directly with version 6 scores.

The same iPad Pro, two different scores

In AppleInsider's test on one M4 tablet, Geekbench 7 came in lower on both single-core and multi-core. The chart above puts the drop at about 14% for single-core and 5% for multi-core.

The tablet itself did not slow down between runs. Geekbench changed what it measures and how it computes the score, which is why AppleInsider says the two versions cannot share a performance chart.

Larger data sets make each test harder

According to Primate Labs, tasks have become more demanding and files larger since Geekbench 6, so its workloads now process more, and larger, files. The stated goal is to push the device harder. File Compression now works on a larger, more varied mix of source code, object code and text documents.

PDF Viewer opens a broader range of files, from park maps to technical documents and academic papers, and the developer and image processing workloads use more assets and different image formats. Neither Primate Labs nor AppleInsider breaks down how much of the iPad's drop comes from the heavier data versus the other changes.

Compare devices only with scores from one version

Use scores from the same Geekbench version when comparing two devices. AppleInsider suggests keeping older results in Geekbench 6 charts, while version 7 needs its own baseline as results accumulate. Mixing the two in one chart would show a gap even between identical hardware, as the iPad Pro test shows.

Before comparing a score found online with your own, check which version produced it. AppleInsider also warns that no benchmark can reproduce the performance of every app or workflow, so a Geekbench 7 score is one data point on a device, not a complete verdict.