- NATURAL 20
- Posts
- Microsoft’s Own AI Models Move Into Bing and PowerPoint
Microsoft’s Own AI Models Move Into Bing and PowerPoint
Wes Roth
July 24, 2026
In partnership with
Try the AI that knows your customers. No commitment.
Most platform evaluations start with a demo request and end three weeks later in a conference room. This one takes 15 minutes and puts you directly inside Gladly's interface — navigating it on your own terms.
See how AI surfaces real-time customer context before a conversation starts. Watch how a single conversation thread pulls in purchase history, channel history, and account details without a handoff.
No installation. No commitment. Start the interactive demo and see the platform for yourself.
Today:
Microsoft launches MAI-Image-2.5-Pro and MAI-Voice-2-Flash
ChatGPT Health connects medical records and Apple Health
Grok 4.5 reaches web, X, iOS and Android
Claude Voice adds advanced models and connected tools
DeepSeek says AGI matters more than short-term profit
Microsoft’s Own AI Models Move Into Bing and PowerPoint
Microsoft is no longer relying entirely on outside companies to power its biggest AI products.
Its newest image and voice models are already moving into Bing, PowerPoint, OneDrive, Dynamics 365, and Azure. At the same time, OpenAI is giving ChatGPT access to personal health information, while xAI is putting Grok 4.5 on nearly every major consumer platform.
Microsoft has introduced MAI-Image-2.5-Pro and MAI-Voice-2-Flash, two internally developed models designed for image creation, editing, speech generation, and real-time voice applications.
MAI-Image-2.5-Pro is Microsoft’s highest-quality image model so far. It can create detailed visuals, edit existing images, and place readable text inside generated designs. Microsoft has made it the default model in Bing Image Creator, meaning Bing’s image-generation system is now powered entirely by Microsoft’s own technology.
The model is also powering image-to-image features in PowerPoint. Microsoft reports that it can reduce the graphics-processing cost of those tasks by as much as 84% compared with GPT-Image-2. In OneDrive, the model increased image save rates by 26%, cut high-end response times by approximately 25%, and delivered around 2.5 times greater efficiency.
MAI-Voice-2-Flash is built for faster and less expensive voice generation. It operates twice as fast as Microsoft’s previous voice model while costing 32% less. Dynamics 365 Contact Center is already using the technology, where Microsoft says it reduced graphics-processing costs by as much as 89%. The model is also being integrated with Azure Voice Live.
Both models are available in public preview. Voice generation costs $15 per million characters, while image pricing depends on the amount of text, source-image data, and generated-image data processed.
OpenAI is rolling out a dedicated Health experience in ChatGPT that can connect with Apple Health and supported medical-record providers.
Once users give permission, ChatGPT can compare test results over time, explain changes in health information, summarize medical records, and relate sleep, activity, exercise, and other personal data to a user’s questions. OpenAI says more than 300 million people already ask ChatGPT health-related questions each week.
The feature can also bring relevant health context into other ChatGPT conversations. For example, a discussion about an exercise routine could use recent activity information without requiring the user to upload the same records again. Users remain in control of which sources are connected and when that information can be accessed.
OpenAI says connected medical records, Apple Health data, and conversations that use this information will not be used to train its foundation models or target advertising. The company also stresses that the feature is intended to support professional medical care—not replace doctors or provide emergency treatment.
ChatGPT Health is beginning to roll out to logged-in users aged 18 and older in the United States. It is available on the web and iOS for Free, Go, Plus, and Pro plans, but it is not currently available inside Codex.
xAI has expanded Grok 4.5 across Grok’s website, X, iOS, and Android, making the model available through nearly every major consumer platform operated by the company.
The release is designed to follow instructions more accurately, maintain longer conversations, reason more quickly, and provide clearer and more dependable answers. It can also read and summarize long PDF documents without requiring users to divide them into smaller sections.
Grok 4.5 can create working spreadsheets, presentations, diagrams, reports, and other structured documents. Instead of providing only instructions, it can generate complete files that users can continue editing.
The model also works inside Microsoft Word, Excel, PowerPoint, and Outlook through Microsoft 365 add-ins. That gives xAI another route into professional workflows without requiring users to leave the applications they already use.
🧠RESEARCH
Researchers introduced SLAI T-Rex, a system for completing full-model post-training of the DeepSeek-V4 family on an Ascend SuperPOD computing cluster. The system reached 34.22% model-compute utilization—a measure of how effectively the available chips were used—and delivered 2.93 times the performance of an open-source baseline.
The researchers also trained a specialized model for operations-research problems using 10,000 supervised examples. It achieved a 71.81% average success rate on its first answer, outperforming GPT-5.4-Mini by 3.98 percentage points and the original DeepSeek-V4-Flash model by 11.27 points.
Most search systems judge each document separately based on how closely it matches a question. This paper argues that the approach overlooks whether a group of documents contains repeated information, conflicting claims, or useful details that complement one another.
The researchers created SetwiseEvalKit, which evaluates complete groups of documents across nine dimensions using approximately 28,000 scoring rules. Even the strongest of 12 tested reranking systems covered no more than 45% of the desired information. Their new Rubric4Setwise method produced better final answers while using fewer documents and fewer search rounds.
ABot-World-0 is an AI video model that generates an interactive world while responding to a user’s keyboard actions. Unlike a normal video generator, it continuously updates the scene based on where the player moves and what they do.
The system can produce 720p video at up to 16 frames per second on a single NVIDIA RTX 5090 desktop graphics card. It begins responding approximately 1.2 seconds after an action and uses about 19 gigabytes of graphics memory. The model also maintains character identity and scene continuity across longer interactions.
📲SOCIAL MEDIA
🗞️MORE NEWS
ChatGPT Voice is coming to the desktop application, allowing users to speak with ChatGPT while it operates a computer or directs several agents working inside ChatGPT Work and Codex. The experience is powered by GPT-Live, which can listen, speak, and coordinate tasks at the same time rather than waiting for a user to finish each command.
Claude’s voice mode now works with the Opus, Sonnet, and Haiku model families and can reach connected services such as Gmail, Slack, Google Calendar, and Canva. Users can ask Claude to summarize messages, reschedule events, prepare a document, or draft an email by voice, with Claude requesting permission before using an outside tool. The beta is available across mobile, desktop, and web in 11 language variants.
Sakana AI has released version 1.1 of Fugu-Ultra, its system for difficult and high-stakes tasks. Fugu presents itself as one model through an OpenAI-compatible interface, but it can assemble and coordinate several expert models behind the scenes. The Ultra version is designed for work such as research-paper reproduction, cybersecurity analysis, patent searches, and competitive data-science problems.
DeepSeek founder Liang Wenfeng says the company will continue prioritizing artificial general intelligence over immediate profits and will probably keep its most advanced models open. He argued that open models and monetization can exist together, while identifying access to computing power—not technical knowledge—as the company’s largest disadvantage compared with American AI laboratories.
DeepSeek is reportedly raising its first outside funding at a valuation of approximately $52 billion. The company is also planning larger computing clusters, although it has not decided whether to develop its own specialized chips.
What'd you think of today's edition? |
Reply