In partnership with
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
Today:
Microsoft Debuts Seven New 'MAI' Models
OpenAI Expands Codex to Non-Developers
Anthropic Confidentially Files for IPO
NVIDIA Cosmos 3 Trains Physical AI
OpenAI Brings Flagship Models to AWS
Microsoft AI has introduced a new family of seven in-house models built from the ground up without relying on distilled data from third-party models. The lab expects a 1,000x increase in compute over the next three years, driving toward what they call "Humanist Superintelligence"—AI that remains subordinate to human goals.
The MAI Lineup & Technical Breakdown
The release introduces a new suite of models optimized for enterprise efficiency and high-volume reasoning, alongside local hardware developments:
MAI-Thinking-1 (Flagship Reasoning): Built on a Mixture-of-Experts (MoE) architecture, this model features 35 billion active parameters and a 256K context window. In independent blind testing, raters preferred it over Anthropic's Claude Sonnet 4.6, and it matches Claude Opus 4.6 on the SWE-Bench Pro coding benchmark.
MAI-Code-1-Flash (Agentic Coding): A highly efficient, 5-billion-parameter coding model deeply integrated into GitHub Copilot and the VS Code stack. It performs comparably to Claude 3 Haiku but at a lower operational cost.
MAI-Image-2.5 (Visual Generation): Microsoft’s flagship model for text-to-image and image editing. It currently ranks #3 on the Arena AI leaderboard for text-to-image and #2 for image-to-image workloads, surpassing Nano Banana Pro. An ultra-efficient "Flash" variant is also available for high-speed tasks.
MAI-Transcribe-1.5 (Speech-to-Text): Built to handle domain-specific terminology natively across 43 languages, achieving state-of-the-art accuracy while operating up to 5x faster than competing models.
MAI-Voice-2 (Text-to-Speech): A natural-sounding speech generator supporting 15 languages, featuring zero-shot voice adaptation from very short audio samples.
Hardware Symbiosis: To support local development, Microsoft announced the Surface RTX Spark Dev Box. Powered by an NVIDIA RTX Spark chip, it delivers 1 petaflop of AI compute and 128GB of unified memory, allowing developers to run models up to 120 billion parameters locally without cloud GPUs.
The models will also be available on Open Router, Fireworks, and Baseten.
Moving beyond standard fine-tuning, Microsoft introduced reinforcement learning environments ("training gyms") that allow models to adapt specifically to a company’s unique workflows. Early testing, including a custom Excel tuning, shows these adapted models can match GPT 5.4 while being up to 10x more efficient.
Microsoft is co-creating a specialized, highly accurate healthcare model with the Mayo Clinic. The model will be owned by the hospital system and eventually offered to other organizations via Azure Foundry.
OpenAI is expanding Codex far beyond its software engineering roots. With non-developers now making up 20% of Codex's 5 million weekly users (and growing three times faster than developers), the platform is pivoting to support everyday business workflows.
Role-Specific Plugins: Codex now features out-of-the-box plugins tailored for distinct roles without requiring any coding. Initial plugins cover Data Analytics, Creative Production, Sales, Product Design, Public Equity Investing, and Investment Banking. These integrate directly with 62 popular apps like Salesforce, Figma, Snowflake, and Canva.
Interactive 'Sites': In a major update for Business and Enterprise users, Codex can now generate, host, and update interactive web apps. Instead of sharing static documents, teams can spin up shareable dashboards, project boards, or scenario planners via a simple URL.
In-Place Annotations: Users can now refine their outputs seamlessly. By highlighting a specific part of a generated site, slide, or document, you can ask Codex to tweak just that section (like updating a chart label or changing a font) without having to regenerate the entire draft.
Anthropic has taken a massive step toward the public markets by confidentially submitting a draft registration statement (Form S-1) to the U.S. Securities and Exchange Commission (SEC).
The IPO Path: A confidential S-1 gives Anthropic the flexibility to launch an Initial Public Offering once the SEC completes its review, though the final decision and timing will depend heavily on market conditions.
Staggering Valuation: While share counts and pricing for the IPO are not yet set, a related announcement on Anthropic's site notes that the company recently raised an astounding $65 billion in Series H funding, catapulting them to a $965 billion post-money valuation.
🧠RESEARCH
"Crafter" is a new system that uses multiple artificial intelligence assistants working together to create excellent scientific charts and diagrams. Unlike older tools that produce flat images you cannot change, this software generates graphics that researchers can easily edit and adjust, saving valuable time during the rigorous research publication process.
Researchers are exploring a new method to build millions of highly personalized artificial intelligence assistants. Instead of creating a completely new massive brain for every user, this approach efficiently attaches tiny customized memory and skill updates onto one giant shared model, drastically lowering computing costs while maintaining very reliable performance.
"OpenWebRL" is a new training method that teaches visual artificial intelligence how to browse the internet by letting it practice on live websites. By learning through trial and error in the real world rather than relying on expensive recorded examples, these programs become much better at navigating complex web pages.
📲SOCIAL MEDIA
🗞️MORE NEWS
NVIDIA Cosmos 3: Teaching AI About the Physical World NVIDIA is releasing a tool that helps artificial intelligence understand how the physical world works, including concepts like gravity, space, and motion. By creating highly realistic virtual environments, robots and smart software can safely practice making decisions and navigating obstacles before they are ever used in real life.
OpenAI Frontier Models and Codex Arrive on AWS OpenAI’s most advanced systems and its popular coding assistant are now officially available on Amazon Web Services. This allows the millions of businesses that already use Amazon's massive network of internet servers to easily plug OpenAI's technology directly into their own products without having to change their existing setups.
Anthropic Expands Project Glasswing Anthropic is growing its initiative designed to give businesses deeper visibility into how artificial intelligence makes its decisions. This expansion provides companies with better tools to look inside the software's thought process, ensuring the system stays safe, predictable, and strictly follows company rules.
Perplexity: Rethinking Search as Code Generation Perplexity is testing a new way to search the internet by having their software write short computer programs to gather answers, rather than just matching your search words to websites. This method allows the system to pull data from multiple sources, perform calculations, and deliver highly precise answers to complex questions that a normal search engine cannot handle.
Alibaba Launches Qwen 3.7 Plus Alibaba has introduced a major upgrade to its artificial intelligence system, making it much faster and smarter at processing both text and images. This update is designed to give everyday businesses a powerful, low-cost option for automating routine tasks and analyzing large amounts of information quickly.
