Jin Daily AI Trivia - Microsoft AI launched 7 new AI models
Microsoft just dropped a bunch of new AI models, but honestly, most of the text and coding models feel very “enterprise checkbox.” The audio models? Different story.
MAI-Thinking-1 - “Don’t think one”
1T-A35B with a 256K context window. Sounds powerful on paper, but the Terminal-Bench result looks pretty bad, which means tool calling and agentic coding are probably not its strength. So unless this is somehow cheaper than DeepSeek V4, which is very unlikely, it feels pretty pointless for general users. Most likely, this is mainly for enterprises that already love the Microsoft stack.
MAI-Code-1-Flash - “Don’t code one”
This is a 5B active-parameter distilled version of a bigger model that Microsoft did not release, probably based on MAI-Thinking-1. But with a Terminal-Bench 2.0 score of only 54.8, it still does not look that competitive. GLM-5.1 and DeepSeek V4 are already above 68. So yeah, another code model that probably should not be your main coding model.
MAI-Image-2.5 / Flash
Microsoft claims better price and performance, but the comparison seems to be against older competitor models. Google Gemini 3.1 Flash Image still beats MAI-Image-2.5 Flash on pricing, so this feels more like a “good enough inside Microsoft ecosystem” model instead of a real market killer.
MAI-Voice-2 / Flash - This is the GOAT
This is the one that actually impressed me. Mixed-language speech in one sentence sounds extremely smooth. Chinese + English, Hindi + English, and other bilingual switching feels very natural. MAI-Voice-2 could kill a lot of paid TTS services out there, especially since it also supports emotion tags.
MAI-Transcribe-1.5
Much language. Much speed. Very wow. 43 languages supported. 1 hour of audio transcribed in around 15 seconds. Microsoft claims it is 5x faster and better than the next SOTA model. This is huge for meeting transcription, podcast processing, call center analytics, and multilingual AI agents.
Don’t need to trust what I say, try it yourself. https://playground.microsoft.ai/chat?model=duo-ai&duo-mode=conversation
Hope you learn something new today!! See ya !!

Related trivia
- AI Trivia · May 6, 2026 Jin Daily AI Trivia : ChatGPT Free Base Model is now GPT-5.5 Instant
- AI Trivia · August 15, 2025 Jin Daily AI Trivia: Did GPT-5 Train on Pirated Drama Subtitles?
- AI Trivia · April 17, 2025 Jin’s Daily AI Trivia: Which OpenAI Model Is This?Today, OpenAI unveiled another “O-series” model—right on the heels of yesterday’s announcement of base model GPT‑4.1 (4.1, 4.1‑mini, and 4.1‑nano). The new lineup now includes:o3o3‑pro (coming soon)o4‑mini‑higho4‑miniConfused by all these names? Let me break down the evolution of the GPT and O-series models for you:GPT‑2 (2019)OpenAI’s first public release (1.5 B parameters).GPT‑3 (2021)Sparked the AI hype.Davinci model: 175 B parameters.Underpins the original ChatGPT.GPT‑3.5 (2022)Introduced the turbo variant and basic browsing.Marked the start of OpenAI’s subscription model.GPT‑4 (2023)Estimated at 1.76 T parameters.First multimodal model (text + image inputs).Free ChatGPT users access GPT‑4‑turbo (text-only).GPT‑4o (2024)“O” stands for “omni.”Real-time text, image, and audio input/output.Free users on image/audio get GPT‑4o‑mini; Plus users use GPT‑4o‑mini as their base.GPT‑4o Imagen (March 2025)Native image generation—an alternative to DALL·E 3.Powers the new “Ghibli-style” image output.GPT‑4.5 (2025)OpenAI’s largest, most expensive model yet.Available only to Plus and Pro subscribers.GPT‑4.1 (2025)A leaner, more cost-effective spin‑down of GPT‑4.5.Replaces GPT‑4o in the lineup.OpenAI‑o1 (2024)First “thinking” model with the “O” prefix.o1‑pro is the priciest, at $150/600 USD per million tokens.OpenAI‑o3 (2025)More efficient successor to o1.o3‑mini launched in January 2025 for Plus users.“o2” was skipped due to trademark issues.Full o3‑pro model to follow later this year.OpenAI‑o4 (2025)Released alongside o3 today.Currently available only as o4‑mini and o4‑mini‑high.Includes basic image-transform tools (crop/rotate) for better analysis.OpenAI rushed to launch o4 to compete with Gemini 2.5 Pro and Qwen. (Not even have a full o4 model)Initial Benchmarking:o4‑mini-high trails Gemini 2.5 Pro by a slim margin—but it’s significantly cheaper, especially for large-token contexts. If you need robust multimodal performance on a budget, o4‑mini is your go‑to.