- Horizon AI
- Posts
- Microsoft Launches Its First In-House AI Models 👀
Microsoft Launches Its First In-House AI Models 👀
Extract anything from a real photo and generate it in 3D
Gina Acosta
August 29, 2025
Welcome to another edition of Horizon AI,
Amid growing rumors of a strained relationship with OpenAI, Microsoft is debuting two powerful new models.
Let’s get into it!
Read Time: 4.5 min
Here's what's new today in the Horizon AI
Microsoft Debuts Two In-House AI Models
OpenAI Releases GPT-Realtime for Developers
AI Tutorial: Extract anything from a real photo and generate it in 3D
AI Tools to check out
AI Findings/Resources
The Latest in AI and Tech 💡
AI News
MICROSOFT
Microsoft Debuts Two In-House AI Models
Microsoft’s AI team introduced two homegrown models: MAI-Voice-1 for fast speech generation and MAI-1-preview, a powerful text model that specializes in following instructions and providing helpful responses to everyday queries.
Details:
MAI-Voice-1 can currently generate about a minute of audio in under one second on a single GPU, making it one of the most efficient speech systems available today.
It already powers some of Copilot’s voice features, like Copilot Daily and Podcasts, and users can also try it in Copilot Labs with custom voices and styles.
MAI-1-preview, which was trained on roughly 15,000 Nvidia H100 GPUs, is being rolled out for "certain text use cases" in Copilot, which still relies heavily on OpenAI’s language models today.
The company is also publicly testing MAI-1-preview on the AI benchmarking platform LMArena and is making it available via API to select “trusted testers.”
Microsoft has invested billions in OpenAI and long relied on its technology to power its AI offerings. With these new models, the company aims to set itself apart and show it is firmly in the AI race, as its relationship with OpenAI shifts from collaboration to competition.
TOGETHER WITH BELAY
Turn Uncertainty into Opportunity with BELAY
Budgets are shrinking. Teams are getting leaner. And the pressure to do more with less has never been greater.
Layoffs. Cost cuts. Uncertainty. You’re feeling it — and so is your business. But surviving a downturn isn’t about cutting corners. It’s about working smarter. With flexible, scalable support, you can maintain momentum and protect your budget — without committing to full-time hires. That’s where BELAY comes in. Our U.S.-based Executive Assistants, Accounting Professionals, and Marketing Assistants deliver the expertise you need — scaled exactly to fit your needs, your budget, and your future. Because the truth is, when you’re stuck organizing your inbox or filing receipts, you’re losing more than time. You’re losing focus, energy, and strategic clarity. And in a downturn, those aren’t luxuries; they’re survival skills. The more you’re pulled into the weeds, the less you’re able to lead your business where it needs to go. Let BELAY help you do what matters most: lead boldly through uncertainty.
Stay lean. Stay focused. Stay ahead with BELAY.
OPENAI
OpenAI Releases GPT-Realtime for Developers
OpenAI moved its realtime API out of beta, giving developers a faster, more natural way to build voice assistants. The new gpt-realtime model speaks and listens directly, skips text conversion, and now reacts to tone and context in the moment.
Details:
The gpt-realtime model handles speech end-to-end, which cuts latency, improves natural tone, and follows complex instructions more reliably.
It can pick up nonverbal cues like laughter, switch languages mid-sentence, and adjust voice style, with two new voices (Cedar and Marin) plus upgrades to existing ones.
Benchmarks show solid gains, including 82.8% on Big Bench Audio, 30.5% on MultiChallenge, and 66.5% on ComplexFuncBench.
Tool use is more dependable, with better function calling, reusable prompts, SIP and remote MCP support, and added image input for reading screenshots or answering visual questions.
Pricing also dropped by 20%, now at $32 per million audio input tokens and $64 per million output tokens, with new options to set token limits, trim multi-turn conversation.
The AI voice market for enterprises is growing more competitive. OpenAI is betting on better instruction-following and voices “that sound more natural and expressive” to stand out.
AI Tutorial
Extract anything from a real photo and generate it in 3D
Go to Gemini and select the 2.5 Flash model.
Upload your image and use this prompt:
"Generate an image of the [element] in this image. White background, 3/4 view. Make it 100% identical to the original and fill almost the entire white canvas."
Download the image and go to Copilot 3D Labs.
Upload your image and generate it in 3D.
AI Tools to check out
🎬 Video Ocean Agent: From one sentence to a polished video, this Agent crafts script, visuals, and voiceover for you in minutes.
🌍 Mirage 2: A real-time, general-domain generative world engine you can play online
📹 Tella: All-in-one screen recorder, to create incredible product demos, tutorials, courses, for Mac & Windows.
🦾 Deforge: Build AI agents visually, no code required.
🤖 Portia: Build AI agents you can trust in regulated environments.
AI Findings/Resources
👀 Older developers more likely to code with AI tools
👉 How to prompt Gemini 2.5 Flash Image Generation for the best results
The latest in AI and Tech
The model is designed for agent-based programming and is described as fast and cost-effective.
It’s available for free for a limited time through launch partners such as GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, opencode, and Windsurf.
The company announced it will use new chat transcripts and coding sessions to train Claude unless users opt out. It will also extend data retention to five years.
All users on Claude Free, Pro, and Max must decide whether to accept the new terms by September 28, while Claude Gov, Work, Education, and API customers are excluded.
If you accept the terms but later change your mind, you can revoke Claude's access via Settings > Privacy > Help improve Claude.
These include a new IDE extension, the ability to move tasks easily between the cloud and your local environment, integrated code reviews in GitHub, and a revamped Codex CLI.
With the new credits, users can use Google’s Flow tool to generate either five free Veo 3 Fast AI videos or one standard Veo 3 video per month.
That’s a wrap!
We'd love to hear your thoughts on today's email!Your feedback helps us improve our content |
Not subscribed yet? Sign up here and send it to a colleague or friend!
See you in our next edition!
Gina 👩🏻💻