Apple Supercharges Its Office Suite with OpenAI

Apple Integrates OpenAI Assistants into Its Office Suite

Apple is expanding its Creator Studio suite with AI assistants in Keynote, Pages, Numbers, and Logic Pro, as well as an integrated effects and template library from MotionVFX in Final Cut Pro. The assistants run on a combination of Apple’s proprietary models and models from OpenAI. The update was announced on October 27, 2026, alongside new iMacs and the 14-inch MacBook Pro featuring the M6 chip. Existing subscribers will receive the updates in November at no additional cost. Apple’s head of marketing for apps and iCloud, Brent Chiu-Watson, called it the most significant update since the launch of Creator Studio.

At the center of the video offering is MotionVFX, previously an independent provider of effects and templates whose acquisition by Apple was already signaled back in March. Final Cut Pro for Mac now directly includes thousands of templates, graphics, titles, and transitions. These are accompanied by tools for rotoscoping, tracking, and color grading. The curated Content Hub featuring stock videos, photos, and backgrounds is likewise built straight into the editing software. Images from the timeline can be edited in Pixelmator Pro without re-importing. Additionally, there are more options for keyframe transitions and 23 new languages for transcription. Motion receives dedicated MotionVFX plugins for color grading and 3D camera tracking. Moving forward, Compressor will be able to encode Cinematic mode videos from iPhone 18 Pro and iPhone Duo.

Pixelmator Pro introduces Generate Fill: users highlight an image area and describe the desired change in natural language. According to Apple, the tool automatically matches lighting, perspective, and texture to the surroundings. The generated elements can be inserted as a new layer or merged with existing layers, including automatic masking.

Logic Pro receives a Liquid Glass-style redesign and an assistant in beta. According to Apple, it can create a vocal track with a single prompt, start new songs, and revise ongoing projects. A newly designed browser brings together loops, instruments, and effect presets, while integrating a Splice library. An on-device analysis called “Music Understanding” automatically classifies and tags newly added sounds. The Stem Splitter now also separates drum tracks without requiring original multitrack recordings. This is complemented by a Producer Pack from salsa producer Sergio George and noise suppression for voice memos from supported iPhones.

In Keynote, Pages, and Numbers, the assistant understands the context of the respective document. In Keynote, it designs, animates, and refines slides; in Pages, it drafts and structures copy. In Numbers, it calculates and visualizes data, suggests chart types, and builds spreadsheets based on general world knowledge. → Apple, The Verge, 9to5Mac, Thurrott, AppleInsider, MacRumors

Synthszr Take: As an independent vendor, MotionVFX sold exactly what subscribers will receive for free starting in November. For template marketplaces, that is the less damaging half of the news. Far more dangerous is the assistant in Keynote that designs and animates slides entirely on its own. A template was always a purchased starting point, and that starting point is now generated by a prompt right inside the document. Consequently, third-party providers lose the exact touchpoint where customers previously turned to their browsers. What remains are niches with distinct personal aesthetics that Apple’s Content Hub does not curate.

With Genie 4, Google DeepMind Now Builds AI-Generated Worlds in Real Time

On October 27, 2026, Google DeepMind unveiled Genie 4, a World Model that generates interactive environments in real time. A simulation can now run for up to two minutes; its predecessor Genie 3 capped out after one minute. The model renders the world frame by frame as the user moves through it. It requires no pre-rendered environments or programmed interactive elements. Users describe a world using a text prompt and navigate through it with their keyboard.

New is the ability to steer the environment in natural language during an ongoing simulation. Users can change the weather, move the camera, or insert new characters and objects. Gemini adapts every input to the existing scene and maintains continuity between prompts. According to Google, Genie 4 is the company’s first real-time world model to generate audio natively. This includes ambient sound, music, and dialogue to which characters react in real time. In terms of physics and human motion, DeepMind claims significant progress, including aerodynamics, collisions, and naturally behaving characters.

Technically, Genie 4 relies on several models from the Gemini family. Gemini Omni is designed to align the appearance and behavior of the simulation with the physical world. Gemini 3.8 Flash operates as an accompanying agent in the loop. It continuously observes the scene, breaks down objectives into individual steps, and checks incoming frames to verify whether each step succeeded. If the user intervenes, the model adapts its plan. In parallel, DeepMind is experimenting with combining Genie 4 and Gemini 4 Argon.

Google openly acknowledges the limitations. In longer simulations, details and objects can shift or distort over time. According to DeepMind research engineer Hang Qi, even when fully utilizing the two minutes, there is no guarantee that a scene will stay consistent from start to finish.

Google highlights film pre-visualization, game prototyping, training environments for robotics and agents, as well as applications in journalism and manufacturing as primary use cases. According to the company, developers from the gaming and film industries are already testing the model. DeepMind frames world simulation as a foundational building block toward artificial general intelligence. The unveiling of Genie 3 last year had put shares of game publishers under pressure. Genie 4 is rolling out as a Research Preview for selected testers. → Google DeepMind, The Verge, Ars Technica, Quartz

Synthszr Take: Every single second in Genie 4 is rendered live, frame by frame, and every second incurs inference costs. Two minutes instead of one effectively doubles the bill per session. That explains why Gemini 3.8 Flash monitors the scene rather than a flagship model. A game studio pays for its assets once; Google pays afresh for every run. As long as that remains true, Genie will find its home in previs and robotics training, where a minute of simulation replaces costly shooting days or hardware test runs. Launching within the Ultra subscription tier fits this logic.

OpenAI Halves Prices and Makes AI Agents Three Times Faster with GPT-6.1 Terra

According to BleepingComputer, OpenAI has released the GPT-6.1 Terra model. AI agents are expected to work three times faster while pricing drops by 50 percent. The version number positions Terra as an update within the GPT-6 series. The focus is on use as an agent—meaning tasks that a model completes autonomously across multiple steps. → BleepingComputer

Synthszr Take: For agents, the price per completed task is what counts, and that multiplies with every intermediate step. Half the price at triple the speed therefore shifts the economics of entire workflows that previously failed on the cost sheet. OpenAI is thus putting pressure on competitors who sell their agents on quality rather than unit costs. Benchmarks are lacking, and the speed claim comes from the vendor itself; companies are primarily buying a price signal. For many procurement departments, that is already enough.

Anthropic spends only 15 percent of revenue on computing power

According to a comparison by Hello China Tech, Anthropic operates with a significantly higher gross margin than Hong Kong-listed Chinese AI labs Zhipu and MiniMax, based on its own IPO prospectus. The analysis is based on half-year figures through late June 2026, which Anthropic submitted to the US Securities and Exchange Commission (SEC) with its S-1 on October 26. Anthropic generated $16.3 billion in revenue in the first half of the year, Zhipu generated around $142 million converted, and MiniMax $116.6 million. According to the company, Anthropic’s cost of revenue consists almost entirely of computing power for customer workloads and stood at $2.4 billion, approximately 15 percent of revenue. Zhipu achieved a gross margin of 26.4 percent, MiniMax 17.9 percent. Anthropic spent 60 percent of its revenue on research and development including training compute, while Zhipu spent 223 percent and MiniMax 255 percent. → Hello China Tech

Synthszr Take: Anthropic spends 15 percent of revenue on customer compute, while MiniMax is left with a gross margin of just 17.9 percent. Procurement departments pay the difference for a vendor the head of compliance signs off on, not for a better benchmark score. This premium holds as long as switching models remains a governance issue. Once Chinese model weights run on trusted hyperscalers, the trust argument shrinks down to brand perception. Then the price per token becomes the bargaining baseline again, and Anthropic’s margin turns into a target.

Amazon considers token consumption as a measure of AI usage 'profoundly flawed'

In an internal memo, Amazon advises its software teams against measuring AI usage solely by token consumption or generated code. According to CNBC, the company describes such metrics as 'profoundly flawed.' The paper had been circulating internally since September and was published on the company’s website on Tuesday. It redefines the role of developers and managers in the AI era. In addition to token consumption, Amazon also lists the adoption rate of AI tools as well as code changes and reviews among problematic metrics because they could drive higher costs and unnecessary complexity. These figures are not discarded entirely: they should be viewed 'with care and judgment' alongside other data, with the benchmark being long-term outcomes. The background is Tokenmaxxing—maximizing AI consumption regardless of return. → CNBC

Synthszr Take: A leaderboard for token consumption is a cost counter with a trophy. Amazon learned that in May when employees simply burned tokens to climb the ranks. Any metric that rewards input drives up the cost per task. The budget pays for every single one of those tokens. Feature velocity, by contrast, counts what actually ships to production. That leaves the key question ahead of the next budget cycle: What does a successful launch cost? Enterprises with adoption rates on their dashboards face the exact same reckoning.

beehiiv opens newsletter data to Claude and ChatGPT via MCP

Newsletter platform beehiiv is offering access via the Model Context Protocol (MCP), allowing language models such as Claude or ChatGPT to connect directly to a beehiiv account. According to the company, access is secured and encompasses content, subscriber data, and performance metrics. Users can query their newsletter data in natural language, generate analyses, and build workflows based on this data. As a use case, beehiiv cites analyzing which subject lines performed best across the ten strongest editions. Another example is a weekly report on revenue and ad performance over the past 30 days. → STACKED MARKETER

Synthszr Take: Figuring out which subscriber segments unsubscribe most frequently used to require a dashboard with predefined filters. Via MCP, publishers formulate the question themselves, and Claude builds an analysis no product manager anticipated. In doing so, beehiiv gives up its own analytics interface as a retention driver: customer lock-in will henceforth come from the data inside the account. It hits third-party providers who make a living selling prettier newsletter charts even harder: their product is now a prompt.

AMD markets business PCs as machines for continuously running AI agents

AMD is promoting its processors for enterprise customers under the new term 'Agentic PC.' According to the manufacturer, such a machine is an ordinary work PC on which humans are meant to collaborate with AI agents. Depending on the configuration, it can also be operated as a dedicated 'Agent Computer' on which autonomous agents run continuously. AMD frames this within an evolutionary sequence: in 2023 and 2024, AI assistants primarily answered questions, summarized meetings, or polished emails. For that, computing power of a few dozen TOPS was sufficient. For 2025 and 2026, the company describes agents that understand context across multiple applications, plan projects, organize inboxes, or draft presentations from notes. → The Deep View

Synthszr Take: AMD has quietly moved the goalposts. While a few dozen TOPS were sufficient for the assistants of 2023 and 2024, agents are now said to primarily need high thread counts and large amounts of local memory. Those exact specs have always been what sells higher-priced configurations. The new label is therefore a lever to lift average selling prices in the commercial segment. Next quarter’s commercial figures will reveal whether IT buyers are willing to pay the premium, especially since most agents currently run in the cloud.

Microsoft offers buyouts to 60,000 employees

According to TechRadar, Microsoft is offering around 60,000 employees the option to voluntarily leave the company with severance packages. The offer applies to staff up to and including the Vice President level. Locations in the US, Canada, India, and the UK are affected. At the end of June, Microsoft had around 223,000 employees. The program is therefore aimed at a good quarter of the entire workforce. → TechRadar

Synthszr Take: A voluntary buyout program leaves the choice to employees, and those with the highest market value are the first to sign. These are people who could start at an AI lab or a startup tomorrow. Because the offer extends up to the VP level, product knowledge and customer relationships are also at stake. Microsoft is buying itself a quieter transition than any wave of layoffs. With a good quarter of the workforce receiving the offer, the corporation loses control over who leaves.

Ahead of Israel’s Election, Repurposed Anti-Iran Bots Campaign for Netanyahu, Deepfakes Flood X

Ahead of Tuesday’s Israeli parliamentary election, bots, AI-generated videos, and voter-tracking apps are being widely used in the campaign, according to Al Jazeera. The liberal daily newspaper Haaretz reports that dozens of bot accounts originally set up against the Iranian government are now spreading propaganda for Benjamin Netanyahu and his Likud party, warning against opposition politicians Gadi Eisenkot and Naftali Bennett. Both camps are circulating deepfake videos on X. Some show Netanyahu grappling with his alleged failures prior to October 7, 2023. Others depict the opposition as yielding to Arab parties, who are seen as potential kingmakers. Polls show the Netanyahu bloc and the opposition under Eisenkot neck and neck. Likud intended to use the Elector app on election day to track turnout, a move restricted by the Central Elections Committee on data privacy grounds.

Synthszr Take: A bot network built against Tehran is now campaigning for Likud, and nobody had to develop new technology for it. What’s scarce is X’s moderation capacity. With a dead heat in the polls, a single deepfake that goes viral on Sunday is enough, while fact-checking arrives only after Tuesday. Platforms size their teams for average load, but election campaigns create peak loads. That gap can be exploited deliberately, and according to the Joint List, Likud has the biggest budget in the race.

Anthropic and OpenAI Claim to Have Disproven the Exponential Time Hypothesis Using AI

An internal research model from Anthropic reportedly found a deterministic algorithm that solves 3-SAT—the satisfiability of propositional formulas with three literals per clause—in subexponential time. This would disprove the Exponential Time Hypothesis (ETH). According to the company, the algorithm even counts the satisfying assignments, which is considered a harder problem. Anthropic itself emphasizes that the result is purely asymptotic and says nothing about practically solving formulas. The paper and formalization are hosted on GitHub as a “research artifact,” are not maintained, and do not accept contributions. The author of the paper is listed as “Claude (Anthropic)”.

The proof is formalized in Lean 4, also written by AI models. A short file states what is supposedly proven. The Lean kernel checks the proofs, but according to documentation, no machine verifies whether this file asserts the same thing as the paper. It was therefore deliberately kept concise so that readers can verify the match themselves. An appendix lists the discrepancies between the formal and printed versions. According to the company, the build takes around eight minutes on 24 virtual CPUs and relies only on Lean’s three standard axioms. A second, independently written kernel checks the proofs again. Anthropic claims that Claude Opus 5.5 can independently re-derive a weaker variant of the result from a heavily condensed description via tensor networks.

On October 26, OpenAI likewise reported a disproof of ETH: a deterministic algorithm that counts satisfying assignments of bounded-width CNF formulas in subexponential time and with polynomial space. The company announced two further results. One claims to prove Sogge’s 1991 local smoothing conjecture in all dimensions, thereby resolving the Kakeya conjecture and Stein’s restriction conjecture. Previously, it had only been proven in two dimensions. The third result claims to demonstrate the mathematical existence of certain quantum field theories in four dimensions, including quantum chromodynamics. According to company statements, the ETH proof was generated on October 16 and formalized one day later. Dan Roberts, who leads OpenAI’s Frontiers team, is responsible.

For publication, OpenAI is relying on external “shepherds” who recommended three different strategies. The mathematics result is published directly as model output. For computer science, a working group led by Russell Impagliazzo and Ramamohan Paturi—who co-formulated the ETH in 1999—is reviewing the proof alongside Ryan Williams from MIT, and the material is public. The physics group led by David Gross is withholding the AI artifacts for now and intends to draft its own write-up first, drawing criticism from independent mathematicians. → unite, GitHub, Al Jazeera, The Neuron, Thoughts, OpenAI, The Verge, Gizmodo, blockchain, RuntimeWire

Synthszr Take: An eight-minute build on 24 CPUs proves that the proofs pass in the Lean kernel. According to the documentation, no machine checks whether the concise statement file matches the paper. The critical verification rests with the reader, and Anthropic places it in a repository that does not accept contributions. OpenAI hands the same claim over to Impagliazzo and Paturi, who co-formulated the ETH in 1999. With claims of this magnitude, credibility hinges on a signature—and a kernel doesn’t sign anything.