AI NewsSeptember 26, 2026

Microsoft unveiled its biggest Copilot update yet, rebuilding it as a work OS with a persistent agent that runs independently.
Microsoft's update introduces four features: Home, a central hub blending chats with team tasks; Code, for building apps using plain language; deep integration across Word, Excel, and PowerPoint; and Autopilot, a persistent agent that works on tasks independently. A new Today dashboard surfaces emails, meetings, and tasks without being asked.
Microsoft executives frame this as a shift from raw AI models to practical, everyday products, with a focus on enterprise needs like compliance. The market noticed too, with shares rising over 3% to $516.17 on the news.
Key Capabilities:

• Autopilot agent: Works proactively and independently on long-running tasks, not just on command.
• Office-wide integration: Built directly into Word, Excel, and PowerPoint workflows.
• Today dashboard: Surfaces relevant emails, meetings, and tasks automatically each day.

Lightreel AI launched Jev, a tool that searches Meta's entire ad library to find similar ads and creative inspiration instantly.
Lightreel introduced Jev for Meta ads. Just describe an ad you're picturing, and Jev searches the largest Meta ad library to surface similar companies, matching ad types, and creative inspiration.
The pitch is simple: marketers currently burn hundreds of hours manually scrolling Meta's ad library for competitor research. Jev turns that into a single search query instead.
Key Capabilities:

• Instant ad search: Describe an ad concept and get matching real examples from Meta's library.
• Competitor discovery: Surfaces similar companies running comparable ad campaigns.
• Creative inspiration: Speeds up ideation by showing what's already working in your space.

Exa launched Agent Ultra, a research AI that deploys swarms of agents to dig through thousands of sources at once.
Exa introduced Agent Ultra, calling it a step change in deep research. It orchestrates swarms of agents to perform exhaustive research, build comprehensive lists, and answer questions that require pulling from thousands of sources.
On benchmarks, it beat rivals by wide margins. On DeepSearchQA, a tough multi-step research test, it scored 93.9% versus Perplexity Agent's 89.7% and GPT-6 Astra's 85.3%. It also crushed a company-list-building task, finding 2,451 matching entities versus under 150 for every competitor.
Key Capabilities:

• Agent swarms: Coordinates multiple AI agents at once to cover ground a single agent can't.
• Top research scores: Leads rivals like GPT-6 Astra and Opus 5.5 on deep research benchmarks.
• List-building power: Found over 2,400 matching companies versus under 150 for competitors.

Meituan launched LongCat-2.5-Preview, a massive multimodal AI model built for long, complex tasks across apps and tools.
LongCat-2.5-Preview is now live from Meituan, packing 1.6 trillion parameters with about 48 billion active at once. It comes with a 1 million token context window, letting it hold huge amounts of information in a single conversation.
The model is natively multimodal and designed for long-horizon tasks, meaning it can work through multi-step jobs without losing the thread. It spans everything from terminals and browsers to GUIs, spreadsheets, and design tools.
Key Capabilities:

• Massive context: Handles up to 1 million tokens, ideal for huge documents or codebases.
• Multimodal by design: Works natively across text, visual, and tool-based interfaces.
• Broad tool reach: Operates across terminals, browsers, spreadsheets, and design software.

Sarvam AI launched Saaras V4, a major upgrade to its speech recognition model built for both English and Indian languages.
Sarvam released Saaras V4, its most capable speech-to-text model so far. It delivers stronger accuracy across a much wider range of speech than earlier versions. The model was built with specific attention to noisy audio, regional accents, and mixed-language speech.
That last part matters in India, where conversations often blend Hindi, English, and regional languages in a single sentence. Most speech models choke on that. Saaras V4 is built to handle it natively.
Key Capabilities:

• Wide language coverage: Strong performance across English and multiple Indian languages in one model.
• Noise-resistant: Trained specifically to handle noisy, real-world audio conditions.
• Mixed-language ready: Built to transcribe accents and code-switched speech accurately.