Geekbench includes updated CPU workloads and new Compute workloads that model real-world tasks and applications. Geekbench is a benchmark that reflects what actual users face on their mobile devices and personal computers.

Features

CPU Benchmark

Geekbench 7 measures your processor’s single-core and multi-core power, for everything from checking your email to taking a picture to playing music, or all of it at once. Geekbench 7’s CPU benchmark measures performance in new application areas including video conferencing, streaming, and game physics, so you’ll know how your system handles the tasks you rely on every day.

GPU Benchmark

Test your system’s potential for GPU compute tasks including machine learning, image processing, and video editing with the GPU Benchmark. Test your GPU’s power with support for the OpenCL, Metal, Vulkan, and CUDA APIs. New to Geekbench 7 is an increased focus on Machine Learning, and support for NVIDIA’S CUDA API.

Real-World Tests

Geekbench uses practical, everyday scenarios and datasets to measure performance. Each test is based on tasks found in popular real-world apps and uses realistic data sets, ensuring that your results are relevant and applicable. Geekbench 7’s multi-core tests only run multi-threaded when real applications do, so your scores reflect how software actually behaves.

Cross-Platform

Compare apples and oranges. Or Apples and Samsungs. Designed from the ground-up for cross-platform comparisons, Geekbench 7 allows you to compare system performance across devices, operating systems, and processor architectures. Geekbench 7 supports Android, iOS, macOS, Windows, and Linux.

Geekbench Browser

Upload your results to the Geekbench Browser to share them with others, or to let the world know how fast (or slow) your devices can go! You can track all your results in one place by creating an account, and find them easily from any of your devices.

Benchmark Charts

Verify device performance using the Geekbench Benchmark Charts. Available on the Geekbench Browser, these charts are based on data aggregated from real users in real-world environments. Whether you’re considering a new purchase or are curious about a device’s capabilities, use these charts to make informed decisions.

Stress Tests

Geekbench includes stress tests, which are tests that help determine the stability of your system. Stress tests help you find small problems with your system before they become big problems.

Multicore Aware

Every test in Geekbench is multi-core aware. This allows Geekbench to show you the true potential of your system. Whether you’re running Geekbench on a dual-core phone or a 32-core server, Geekbench is able to measure the performance of all the cores in your system.

Here’s what the different numbers mean:

Battery Runtime is the battery test runtime. If the test started with the battery completely charged and ended with the battery completely discharged then the test runtime is also the battery lifetime.

Battery Score is a combination of the runtime and the work completed during the battery test. If two phones have the same runtime but different scores, then the phone with the higher score completed more work. As with Geekbench scores, higher battery scores are better.

Battery Level is the battery level at the start and the end of the test.

What’s New

Geekbench 7, the latest version of Primate Labs’ cross-platform benchmark, has arrived and features both new and improved workloads to measure the performance of your CPUs and GPUs.

New Media Workloads

Geekbench 7 includes new media workloads that measure how well your CPU handles audio and video encoding, decoding, and processing. The workloads model the tasks behind video conferencing, screen sharing, and everyday content consumption. These new workloads:

• Encode screen-sharing video with the AV1 codec, modeling the screen-sharing features in video conferencing apps.
• Compress music and spoken-word audio with the Opus codec, modeling voice memo and podcast apps.
• Decode video and audio while generating live captions with the Whisper speech recognition model, modeling video playback with automatic subtitles enabled.

Alongside the new media workloads, Geekbench 7 adds a Game Physics workload built on the Jolt Physics engine used in popular video games, expands the Photo Editor workload with a richer set of real-world edits, and updates the Photo Library workload to support importing and processing modern image formats such as JPEG XL and DNG.

A Smarter Multi-Core Benchmark

The multi-core benchmark in Geekbench 7 has been redesigned to better reflect how real applications behave. Not every task in the real world is multi-threaded, and pretending otherwise distorts scores without telling you anything useful about your device.

In Geekbench 7, a workload only runs in multi-threaded mode if the task it models actually runs multi-threaded in real applications. For example, the HTML5 Browser test isn’t included in the multi-threaded suite because web browsers are single-threaded (or lightly threaded).

The result is a multi-core score that’s a more accurate, more useful, and more relevant measure of how your device performs the work you actually do.

A Refreshed GPU Benchmark

The GPU benchmark has a new focus on the machine learning and content creation applications that increasingly define GPU performance. For machine learning, the GPU benchmark includes workloads that:

• Track faces and apply real-time filter effects to video, modeling the face filters in social media apps.
• Upscale images with machine learning, modeling super resolution features in content creation apps.
• Blur backgrounds in video conferencing streams, modeling virtual background features in video conferencing apps.

Geekbench 7 also introduces new GPU image editing and synthesis workloads, including RAW image processing, LUT-based video color grading, path tracing, and fluid simulation.

And by popular demand, CUDA joins OpenCL, Vulkan, and Metal as a supported API in the GPU Benchmark. You can now measure your NVIDIA GPU using the API that powers its most demanding applications.

Larger Data Sets

Since the release of Geekbench 6, the tasks people perform have become more strenuous, and the data sets they use have grown larger and more demanding. To reflect this, we’ve updated the data sets the workloads process so they’re more challenging for your device and better reflect the files people work with today. This includes:

• More (and more varied) archives in the File Compression workload, spanning source code, object code, and text documents.
• More (and more varied) documents in the PDF Viewer workload, from park maps to technical documents to academic papers.
• More assets and different image formats throughout the developer and image processing workloads.

• Nvidia’s Rubin-generation AI servers run coolant that is often hotter than a hot tub as it enters a closed loop at up to 45°C and leaves at around 55°C, with no performance penalty in tow
• The principle leveraged here is that of a temperature gradient with server chips still running hotter than the coolant, allowing heat transfer to occur seamlessly
• The approach allows Nvidia to roughly conserve 2.6 million gallons per megawatt annually to near zero and could save a 50MW site over $4 million a year in cost

Nvidia‘s Rubin generation is the first of its kind in multiple ways, but the one that sticks out possibly is the fact that it is 100% liquid cooled, implementing it as a core design feature at a platform level with every single chip and networking component covered by a closed loop.

The reference design by Nvidia aims to allow for AI factories that effectively consume zero water by implementing a closed loop circuit that does not leverage evaporative water cooling 99% of the time.

A federal district judge will oversee Apple’s lawsuit accusing OpenAI of using stolen intellectual property to advance its hardware efforts.

In early July, Apple sued OpenAI after alleging that two ex-employees successfully stole intellectual property to enrich OpenAI’s development efforts. Now, on Thursday, a judge has been assigned to the case.

Initially, when the suit was first filed, it was randomly assigned to Magistrate Judge Virginia K. DeMarchi. Now, it seems as though U.S. District Judge Edward Davila will oversee the case.

As 9to5Mac notes, each party was given the choice whether to allow the magistrate judge to sit the case. Apple appears to be the party that declined, instead opting to reassign the case to a district judge.

The case’s previous initial case management conference had been scheduled for October 13. However, it will now need to be rescheduled before Judge Davila.

Maintained intellectual property theft

Apple’s argument against OpenAI hinges on its argument that two previous employees had been stealing intellectual property for quite some time. The two employees are Chang Liu and Tang Yew Tan, the former Vice President of Product Design for iPhone and Apple Watch.

Reportedly, Liu failed to return Apple-issued hardware that was still authenticated to access Apple’s networks. He allegedly told a colleague, Yu-Ting “Alyssa” Peng, still at Apple, that he was planning to access Apple information.

Tan, however, allegedly began emailing himself information about Apple suppliers months before he left to serve as OpenAI’s Chief Hardware Officer. He also allegedly directed candidates to bring unreleased hardware components from Apple to their interviews with OpenAI.

Apple believes this was part of a concerted effort to take and use confidential information. Apple is seeking judgment, an injunction against use and possession of Apple’s intellectual property, a return of Apple’s property, damages, and royalties for use of Apple’s intellectual property.

A malvertising campaign on the Bing search service is pushing a fake Claude desktop app installer hosted on a legitimate Claude.ai domain to deliver the SectopRAT malware.

At least 29 organizations were compromised between July 21-22 during the malicious operation, which researchers call FakeAgent.

The attacker uses a malicious Claude Artifact hosted on Claude’s legitimate domain, which is a common tactic that has been used 

The attackers used a malicious Claude Artifact hosted on Claude’s legitimate domain, a tactic that has been used at the beginning of the year to push macOS malware via ClickFix lures.

Researchers at managed security company Huntress found that the malicious Claude Artifact, downloaded 7,100 times before Anthropic removed it, directed visitors to websites that hosted a fake installer named ClaudeDesktop.exe.

However, the file is a legitimate JetBrains Chromium component that sideloads a malicious DLL (libcef.dll) to deliver the SectopRAT remote access trojan with info-stealing capabilities.

Persistence on the system is achieved through another executable named DockerDesktop.exe, which installs a scheduled task.

Huntress says that the various loaders and staging components used in the infection chain feature anti-analysis mechanisms, including VMProtect packing, shader timing checks, GPU and VRAM checks, and virtual machine (VM) detection.

The SectopRAT malware was recently observed being distributed via CastleLoader campaigns and ClickFix attacks.

The malware uses the EtherHiding technique to retrieve a working command-and-control (C2) address via Ethereum BNB Smart Chain transactions.

SectopRAT, also known as ArechClient2, has been active since 2019 and is an information-stealer with HVNC (Hidden Virtual Network Computing) functionality. It allows remote hands-on operations and real-time interaction with the compromised system.

The malware targets user passwords, credit card data, files, browser logins and cookies, FTP credentials, data from various messaging clients, including Discord and Telegram, Steam, and VPN products.

In an interesting twist, Huntress reports it used Claude Opus 4.8 to assist with shader emulation, cryptographic reconstruction, and .NET code analysis.

Analysis of the decrypted .NET payload (SectopRAT) helped attribute the attacks to SectopRAT operations, or a closely related fork, and opened the path to infrastructure analysis.

At that stage, the researchers found 10 domains registered to the same email address since December 2025, with one of them previously linked to the StealC distribution and seized during Operation Endgame.

Huntress does not have enough evidence to attribute the FakeAgent campaign to a specific, known threat cluster.

Users looking for software should trust official websites and download portals, instead of search results, especially sponsored ones.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Get the whitepaper

For months, AI giants have devised special vetted programs and strict guardrails to limit the use of their models by malicious hackers. But these limits are now hindering the work of legitimate network defenders, as well as that of offensive cybersecurity researchers. 

In June, the U.S. government slapped export control restrictions on Anthropic’s much-hyped AI models Mythos and Fable. The move was prompted at least in part by a report that claimed it was possible to bypass the models’ guardrails designed to prevent users from using them to build and execute malicious cyberattacks.

Regardless of whether the incident was really motivated by fears of a jailbreak, the fact is that Anthropic has repeatedly marketed Mythos as some kind of doomsday cybermachine that can only be given to carefully vetted users, and even then with strict guardrails in place. (The export controls on Fable 5 and Mythos 5 have since been lifted. Fable 5 returned to general access on July 1; Mythos 5 has been reintroduced only to vetted U.S. organizations as part of the government’s review process.)

That kind of gatekeeping isn’t unique to Mythos. Both Anthropic, with its other models, and OpenAI offer cybersecurity researchers programs they can apply to get vetted and — if approved — access models with fewer cybersecurity restrictions: OpenAI’s Trusted Access for Cyber and Anthropic’s Cyber Verification Program. 

These guardrails have been widely criticized, particularly by researchers whose job is to find unknown vulnerabilities in systems and devise ways to exploit them before criminals do.

During a recent appearance on a cybersecurity podcast, Mark Dowd, a well-known security researcher, said that, “it’s not really comfortable to me that these random large companies are making arbitrary decisions about what is safe in security and what’s not.”

Dowd has spent decades finding and selling “zero days” — previously unknown software flaws and the exploits that take advantage of them — to Western governments, rather than report them to the software makers so they get patched. Governments pay a premium for vulnerabilities precisely because they stay open, which is useful for intelligence operations.

Dowd admitted his work may make him biased, but he isn’t alone. Several people who work in offensive cybersecurity — they proactively probe systems for weaknesses — described to TechCrunch how they use AI tools and deal with their guardrails. 

Chris Anley, the chief scientist at security consulting giant NCC Group, said that asking an AI model to try to exploit a bug is a key step in confirming it’s a real vulnerability worth fixing. But if a guardrail prompts the model to refuse to answer the question outright, the guardrail hurts defenders, he said.

“This is where the whole offensive versus defensive and guardrails part comes in, because ‘fix this code’ as a prompt is both an essential mechanism for defense but also a roadmap for finding critical vulnerabilities in the code base,” said Anley. “So at the same time, the same tool is both an offensive tool and a defensive tool, and the two can’t really be unpicked.”

It’s “like a hammer,” he continued. “You can’t build a house without a hammer. It’s definitely a tool but it’s also irreducibly a weapon as well.”

When he and his colleagues run into such a roadblock, they sometimes fall back on open-source AI models that come with no guardrails at all.

Paolo Stagno, the chief technology officer at CrowdFense, a well-known company that develops, acquires, and sells unknown vulnerabilities to government agencies, agreed with Dowd, saying AI companies “essentially treat customers like children who need babysitting” with their vetted programs and guardrails. 

Stagno said he and his colleagues do use frontier models — but only for reverse engineering. They avoid using AI to help find vulnerabilities or build exploits, he said, because feeding that work into a cloud-based model risks leaking sensitive vulnerability data or having it absorbed into future training runs. For that step, he said, they use open source models run locally, as they do not rely on sharing data outside of the model. 

Giuseppe Cali, a security researcher who finds zero-days and develops exploits, said guardrails are not impeding his work. That’s because he doesn’t use AI for offensive work; instead, he uses it for initial reverse engineering, to understand the code he’s analyzing, and to build supporting tools. For that, he said, AI tools can speed up the process and allow him to focus on discovering vulnerabilities. 

“I still want to own the actual bug discovery and weaponization myself and that wouldn’t change if all guardrails were lifted tomorrow,” said Cali. “I am jealous of my bugs, and I like this game too much to let models play it for me.”

One researcher at a smartphone-component manufacturer, who spoke on condition of anonymity because he isn’t authorized to talk to the press, said his employer isn’t part of Anthropic’s CVP program and as a result, its tools are barely useful for finding vulnerabilities because the guardrails are too strict.

“If it catches wind we’re doing anything security related, it just stops and isn’t usable,” the person said. 

Chris Thompson — chief executive of cybersecurity firm RemoteThreat and founder of Offensive AI Con, an offensive security and AI-focused event — said that in his experience using the frontier AI models, the guardrails can be inconsistent and work differently every day. That’s true even inside the looser boundaries of Anthropic and OpenAI’s vetted programs. 

“I think the practical impact is you spend a lot of time negotiating with the model instead of working on the core security program,” said Thompson. “Instead of analyzing a vulnerability and reasoning through the exploitability, you’re trying to find why you’re getting inconsistent results or why are models over-sanitizing the output.” 

As a consequence, researchers rely on or get pushed toward Chinese open-source models like GLM — freely downloadable models that can be run locally with no vetting or usage restrictions — said Thompson.

“You have these responsible researchers that are being pushed away from U.S.-governed systems to foreign-owned systems,” he said. “I think it’s more harmful than good to have these guardrails in place.”

Rather than tightening restrictions further, Thompson called for the AI frontier labs to open up their programs, provide responsible access, and also hold those who abuse their tools accountable. Otherwise, he argued, defenders will lose the AI race.

“There’s this big storm coming. There’s this big wave of attacks that are going to happen at speed and scale like never before,” said Thompson. “But the same security consulting firms and legit researchers that are trying to make a difference are being stifled right now.”

Elon Musk has reignited speculation of a potential Tesla-SpaceX merger.

Tesla has missed recent analyst expectations by a large margin, reporting roughly $1.1bn in adjusted net income in the quarter past. Wall Street had expected the company to rake in around $1.9bn.

The loss is despite the Elon Musk-owned company recording a 26pc growth in revenue to a better-than-expected $28.2bn, with car sales alone bringing in more than $20bn – marking a 23pc year-over-year growth.

Electric vehicle (EV) sales jumped 25pc since last year to a little more than 480,000 units in Q2 2026. The company produced around 451,000 vehicles during the period.

Tesla stocks dropped around 1.3pc at market close on Wednesday (22 July) following the announcement and fell a further 4.2pc in after-hours trading. Company shares have been down nearly 8pc since last month and nearly 17pc in the last six months.

The loss coincides with a drop in share value for Musk’s other company SpaceX, which Tesla has close financial ties with. SpaceX’s filings from earlier in the year showed that the company purchased nearly $700m worth of Tesla’s battery storage products across 2024 and 2025, as well as more than $130m of its Cybertrucks in 2025.

Meanwhile, xAI – now owned under SpaceX – purchased $292m in Tesla battery solutions by April this year, and more than $400m last year. The two companies are also collaborating to develop semiconductors as part of Terafab.

Tesla sales bounced back the strongest in Europe, partially led by higher fuel prices, which drove consumers towards EVs.

New registrations for Tesla vehicles soared across the region, according to June figures – more than doubling in France, and seeing a 39pc rise in Denmark and a 56pc jump in Sweden.

Meanwhile, the company suffered on its home turf in the US after the government cut federal tax credits for EVs and dismantled rules that encouraged their production.

Additionally, Chinese competitors such as BYD, Nio and Xiaomi are also encroaching on Tesla’s EV market share with their more affordable yet high-tech options.

Tesla is attempting to diversify its revenue streams away from EVs (which makes up a majority of its earnings) to autonomous taxis and AI-powered humanoid robots.

“The company reports that paying customers have travelled 2.5m miles in Tesla’s robotaxis and that 380,000 of those miles have been unsupervised, with no safety monitor in the vehicle,” said Forrester VP and principal analyst Paul Miller.

“Those unsupervised miles are rising, but they’re currently a fraction of the 220m miles reported by competitor Waymo back in March.”

Tesla more than doubled its capital spending compared to Q2 last year to fund the diversification push, marking a $1.1bn negative cash flow caused by a capex increase of about $3.3bn.

Musk, meanwhile, told investors that the company aims to spend more than $25bn this year – nearly triple the $8.5bn it spent in 2025. Big Tech heavyweights are expected to commit several hundred billion dollars in capex this year alone to build out their AI ambitions.

Musk also reignited speculation of a possible Tesla-SpaceX merger at the earnings call yesterday.

“As you can tell from the many collaborations on so many fronts with SpaceX, there’s more and more overlap,” Musk said.

“We can’t ​talk about, you know, combining companies and that kind of thing on an earnings call. It’s ​got to be done with the appropriate process.”

A merger could ease matters for the otherwise struggling Tesla, which is already making a pivot closer to SpaceX with its push into AI.

SpaceX president and chief operating officer Gwynne Shotwell told CNBC in June that a combined business “might ⁠make Elon’s life a little easier” by potentially simplifying Musk’s trillion-dollar corporate empire.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

First look: Brooklyn-based consumer electronics company Light has launched a minimalist flip phone designed to cut down on screen time without compromising on essential communication. The $299 Light Flip features a clamshell design reminiscent of the Motorola Razr devices from the turn of the century, and offers a barebones software experience without social media, web browsing, media streaming, and shopping apps.

The Light Flip features a 2.8-inch non-touch OLED display, 6GB of RAM, 128GB of storage, a 50MP rear camera, and a replaceable 1,800mAh battery. It sports an old-school T9 keypad, which used to be standard on candybar feature phones from Nokia, Motorola, and Sony Ericsson in the early 2000s.

Connectivity options include 5G, Wi-Fi, Bluetooth, GPS, and a 3.5mm headphone socket. It is available in six colors and comes with a lanyard notch, adding to its nostalgic appeal.

The Light Flip runs on LightOS, a minimalist Android skin that offers only essential apps and services, such as calls, texts, camera, calendar, music player, and an alarm clock.

As part of its core philosophy of “going light,” the device deliberately skips social media apps and web browsers, encouraging people to use their phone as little as possible. It also supports Light’s digital detox program ‘Flip Your Life,’ which aims to help users adapt to a less connected lifestyle.

Talking to ZDNet, Light co-founders Joe Hollier and Kaiwei Tang said that contrary to popular perception, the target demographic for the Flip are 18- to 30-year-olds who are tired of the 24×7 connected lifestyle. The company believes that dumb phones could make a comeback, as its internal research suggests that 52% of millennial and Gen Z consumers would consider using basic feature phones to cut back on social media consumption and escape the endless cycle of doomscrolling.

The Light Flip is now up for preorder with a $39 deposit but it’s not expected to ship until April 2027.

It will be sold unlocked for $299 from the company’s official website, and customers can also get it on a two-year, $39-a-month service plan directly through Light. Whether it will land on major carriers like Verizon, AT&T, or T-Mobile and become available with a contract remains to be seen.

Microsoft AI released two new in-house models into public preview on Wednesday — MAI-Image-2.5-Pro, its highest-fidelity image generator to date, and MAI-Voice-2-Flash, a speech model built for high-volume enterprise workloads — while publishing production data that amounts to the company’s most aggressive argument yet that it can power its own products without leaning on OpenAI’s frontier models.

The announcement, made by Microsoft AI’s Superintelligence team, lands roughly a year after the company committed to building purpose-built models internally, and it arrives with an unusual level of specificity about where those models now run: Bing, PowerPoint, OneDrive, Dynamics 365, Excel, GitHub Copilot, and Azure. The message to enterprise buyers — and, implicitly, to OpenAI — is that Microsoft’s homegrown models are no longer research projects. They are production infrastructure serving millions of users.

“Each of these enhancements is a step toward the same goal: Microsoft products, powered by Microsoft models,” the company wrote in its announcement blog.

How MAI-Image-2.5-Pro and MAI-Voice-2-Flash stake out opposite ends of the AI cost curve

The two new releases occupy opposite ends of what Microsoft calls the quality-speed-cost curve, and the positioning is deliberate. MAI-Image-2.5-Pro targets the premium tier: hero imagery, detailed editing, and precise in-image text rendering — the last of which has long been a notorious weak spot for image generation models. Microsoft priced the model at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens. The base MAI-Image-2.5 model recently launched at No. 2 for image editing on Arena, the community leaderboard that has become a de facto scoreboard for generative media.

The creative industry appears to be taking notice. Rob Reilly, global chief creative officer at advertising giant WPP, called the Pro model “a strong leap forward for GenMedia tools” in a statement included in Microsoft’s announcement, adding that “Microsoft has firmly established itself among the leaders in generative AI.”

MAI-Voice-2-Flash goes the other direction. First previewed at Microsoft’s Build conference, Flash runs twice as fast as MAI-Voice-2 and costs 32% less, priced at $15 per million characters. It is designed for the unglamorous but enormous market of high-volume voice — call centers, voice agents, and real-time speech applications where latency and cost-per-call matter more than marginal gains in expressiveness. Together, the two models reflect a strategy of building families of models rather than a single flagship, because, as the company put it, a creative studio chasing maximum fidelity has very different needs from a customer service operation handling millions of calls a day.

Microsoft’s production metrics show in-house models cutting GPU costs by up to 89%

The model launches are arguably less newsworthy than the deployment metrics Microsoft attached to them — numbers that read like a systematic case for swapping out third-party frontier models across its product portfolio. 

Bing Image Creator now runs entirely on MAI-Image-2.5, end to end, marking the first time the consumer image tool is fully in-house. In PowerPoint, Microsoft says MAI-Image-2.5 reduces GPU costs by up to 84% compared with GPT-Image-2, OpenAI’s image model. In OneDrive, where MAI-Image-2.5 is now the default for key image-editing scenarios, the company reports a 26% increase in save rates, roughly 25% lower P95 latency, and 2.5 times greater efficiency under medium-utilization production workloads.

On the voice side, MAI-Voice-2-Flash now powers Dynamics 365 Contact Center — the platform used by customers including T-Mobile and EasyJet — where Microsoft claims GPU cost reductions of up to 89%. The model is also integrated into Azure Voice Live for developers building speech-to-speech agents.

Perhaps the most consequential deployment sits in healthcare. Microsoft’s Dragon Copilot, used by 170,000 medical providers and responsible for processing 28 million patient encounters last quarter, now runs on MAI-Transcribe-1.5 for its multilingual workflow across 58 languages. Microsoft says internal evaluations show a 50% relative reduction in both transcription and language-identification error rates across most languages — a meaningful claim in a domain where transcription errors can propagate directly into clinical notes.

Inside the ‘hill-climbing’ strategy that lets small models beat GPT-5.6 in Excel

In a companion post published the same day, Microsoft detailed the methodology behind these results — what it calls its “hill-climbing machine,” an integrated flywheel of data, models, and the product “harness” that surrounds them.

The clearest example is MAI-Code-1-Flash, the lightweight coding model launched in GitHub Copilot in June. Microsoft says the model achieves an approximately 10% higher code accept rate than GPT-5.4 Mini and Claude Haiku 4.5 in VS Code, while using 10% fewer median tokens. Developer retention tells a similar story: users were 6% more likely to return across multiple days than with GPT-5.4 Mini, and 11% more likely than with Claude Haiku 4.5.

Then Microsoft did something more interesting. It took the MAI-Code-1-Flash checkpoint and further trained it inside an Excel reinforcement learning environment, teaching a coding model the tools and workflows of spreadsheet knowledge work. The result, according to production user feedback, is a model on par with GPT-5.6 for the most common Excel tasks — while being small enough to run on Nvidia’s older H100 and even A100 GPUs rather than requiring the latest-generation accelerators.

That hardware detail deserves emphasis. Every major AI company is fighting for allocation of cutting-edge chips, and a model that delivers frontier-adjacent quality on two-generation-old silicon fundamentally changes the deployment economics. It also frees the newest hardware — including Microsoft’s now-operational GB200 cluster — for training rather than serving.

Satya Nadella’s ‘frontier diffusion’ manifesto redraws the OpenAI relationship

Microsoft CEO Satya Nadella framed the announcements in a lengthy post on X titled “Frontier Diffusion & Control,” which functions as something close to a strategic manifesto. “We can now take saturated frontier capabilities and deliver them at scale and at lower cost through models optimized for high-usage products, while continuing to use frontier models for frontier needs,” Nadella wrote, adding that Microsoft is “beginning to route traffic across our first-party surfaces to MAI whenever our models match or outperform frontier alternatives.”

Translated from executive prose: capabilities that were state-of-the-art a year ago are now table stakes, and Microsoft believes it can replicate them cheaply for the specific, repetitive tasks that dominate real product usage. Why pay frontier prices for a frontier model when a user just wants to reformat a spreadsheet column?

Nadella was careful to note that “frontier models from OpenAI and Anthropic are part of the orchestration system alongside MAI” — but he also articulated a pointed principle of model independence, arguing that a company’s evaluations “should continue to hill climb even when any given model has been removed.” 

“Keeping the harness, memory, context, and skills outside the model, he argued, is what gives Microsoft control. The subtext is hard to miss. Reuters reported in April that Microsoft’s exclusive license to OpenAI’s technology had been revised into a non-exclusive arrangement, and The Information reported last September that Microsoft had begun incorporating Anthropic models into some products. Wednesday’s announcement completes the triangle: Microsoft as orchestrator, with its partners’ frontier models as interchangeable components and its own models absorbing an ever-larger share of routine traffic.”

Developers cheer cheaper task-specific models while skeptics question Microsoft’s track record

The response online captured both the appeal and the skepticism surrounding the strategy. “I love when people use small models for niche tasks,” wrote one X user, @mavihsk, responding to Nadella’s post. “Why do I have to use the all-knowing model just to change my field in Excel?” Another user, @nabu_lines, distilled the pitch neatly: “cost and performance both improve when you stop overusing the biggest model.”

Others were less charitable about Microsoft’s execution track record. “Microsoft is the worst when it comes to listening to user feedback,” wrote designer @designedbyabin, arguing the company “will lose the AI race because they repeatedly failed to understand user needs.” And one user, @tokenoverflow, offered a drier critique of the model-independence pitch: “i want it keep hill climbing after removing microsoft.”

The skeptics raise a fair point. Microsoft’s self-reported metrics — accept rates, save rates, GPU savings — come from its own internal evaluations, not independent benchmarks, and the company chooses which comparisons to publish.

But the strategy’s logic does not depend on any single number. Nadella’s framing that software now has “real marginal cost for the first time” explains why Microsoft is obsessive about tokens, GPUs, and serving costs: when AI features run on every keystroke across a billion-user product portfolio, an 84% GPU cost reduction is not an optimization. It is the difference between a viable business and a money pit.

Why Microsoft is turning its internal AI playbook into an Azure product

The final piece of the strategy is that Microsoft is selling the playbook, not just the models. Nadella explicitly positioned the hill-climbing approach as “a template for every other AI native, SaaS, or Enterprise company,” and Microsoft is packaging the toolchain through Foundry and what it calls Frontier Tuning — letting enterprises train specialized models against their own proprietary evaluations and reinforcement learning environments. That turns Microsoft’s internal cost-cutting exercise into an Azure product, and it gives enterprise customers a reason to run their AI workloads on Microsoft’s cloud even if the models themselves come from elsewhere.

The company’s emphasis on models trained “on clean, traceable, enterprise-grade data, without distillation from third-party models” serves the same commercial end. In an industry facing mounting scrutiny over training data provenance, Microsoft is betting that enterprise buyers — and courts — will care where model capabilities come from. Microsoft says it is now extending the hill-climbing approach to Copilot Chat, Outlook, and PowerPoint, and both new models are available in public preview through Microsoft Foundry and the MAI Playground. “None of this is an endpoint,” the company wrote. “We’re just getting started.”

Seven years ago, Microsoft bet more than $13 billion that OpenAI would build the future of AI. Wednesday’s announcement suggests the company has since learned a cheaper lesson: the future of AI may belong to whoever builds the frontier, but the profits belong to whoever makes it ordinary.