AIMarketCap | AI Trends & Prompts

Claude Code Can Now Test iOS Apps Anthropic has added built-in support for the iOS Simulator directly to Claude Code Desktop. AI can now build an app, launch an iPhone or iPad simulator, and navigate the interface on its own. Claude opens the app, taps buttons, moves between screens, checks what changed after each code edit, and if it finds an issue, it immediately fixes the code and runs the test again. The best part is that the simulator now works right alongside the chat window. You can take control at any moment, then hand it back to Claude whenever you’re ready. In practice, developers no longer have to constantly switch between Xcode, the iOS Simulator, and an AI chat. Claude now handles almost the entire development and testing workflow from a single window.

Alibaba Releases Qwen-Image-3.0 The biggest upgrade over the previous version is the prompt length. The model can now process up to 4,500 prompt tokens in a single request, compared to the 1,000-token limit in Qwen-Image-2.0. This allows it to handle tasks that typically break most image generators: newspaper spreads packed with text, multi-layered infographics, scientific paper pages with formulas and charts, and complex app interfaces—all generated in a single pass without stitching separate outputs together. Alibaba also claims support for 12 languages and more than 20 fonts with native text rendering. While text generation has long been a weak point for image models, Qwen-Image-3.0 promises output quality suitable for real-world use. That said, Alibaba has not released any benchmarks, a technical report, or the model weights. For now, the model is available to try through the chat interface. The best AI on AIMarket 🚀

Alibaba Prepares Qwen3.8 Release According to the company, the model features around 2.4 trillion parameters and ranks just behind Fable 5 in overall capabilities, putting it in direct competition with the world’s most advanced AI models. A preview version, Qwen3.8-Max-Preview, is already available through Alibaba Token Plan, Qoder, and QoderWork, giving early users a chance to test it ahead of the official launch. The full public release is expected to arrive in the near future. The best AI on AIMarket 🚀

Lucy 2.5 Can Edit Video in Real Time Decart has unveiled Lucy 2.5, a new AI model for real-time video generation and editing. It can add, remove, or replace virtually any object on the fly, change the style of an entire scene, and apply visual effects without lengthy processing. The model also delivers noticeably higher image quality and realism while following text prompts with greater accuracy. According to Decart, Lucy 2.5 is well suited for applications such as virtual try-ons in e-commerce, interactive live streams, advertising, and other services where fast creation and editing of video content are essential. The best AI on AIMarket 🚀

Stylish Collage Upload your photo and use the prompt below:

A surreal minimalist collage poster on a solid dark YOUR COLOR background, featuring a torn black-and-white portrait of the person from the uploaded photo placed inside a rough paper frame. Add a [LETTER, SHOPPING LIST, RECEIPT, etc.] behind the portrait, [CIRCLES] across the composition, and a bold headline reading “TITLE” inside a small block in the bottom-right corner. Use [PINK] accents, torn edges, pieces of tape, handwritten fragments, a grainy photocopy texture, an asymmetrical layout, strong negative space, a raw editorial aesthetic, and sharp geometric contrast.

The best AI on AIMarket 🚀

Gemini Introduces Avatar for Nano Banana Gemini has rolled out a new Avatar feature for Nano Banana. You only need to record your face and voice once, and the service will create a personalized avatar. After that, you can use it in any prompt without having to upload new selfies every time. The feature is currently being rolled out gradually to Pro users (and above) and is available only in select countries. It is not yet available in the European Union, the United Kingdom, or Switzerland. The best AI on AIMarket 🚀

ChatGPT Just Upgraded Chat Search ChatGPT’s search has been updated to work across multiple types of content. You can now search not only through your conversations, but also across projects, images, and documents—all from a single search window. The feature is already available on the web, as well as on iOS and Android. The update also introduces handy filters that make it much easier to narrow down your results. When you select a match, ChatGPT jumps directly to the relevant message, project, or file. Previously, finding an old conversation or document often meant scrolling through your chat history for minutes. Now, you can simply type a keyword and find what you’re looking for in seconds. The best AI on AIMarket 🚀

Lightricks has launched LTX Reframe LTX Reframe is a video aspect ratio conversion tool built on top of the LTX-2.3 model. Instead of cropping or stretching your footage, it keeps the original frame intact and generates only the areas outside the existing boundaries, naturally extending the scene and preserving motion in the new format. The result is a video optimized for virtually any platform. It supports 16:9, 9:16, 1:1, as well as 4:5 and 5:4 aspect ratios for social media. It handles videos up to 60 seconds long and outputs native 1080p—without black bars or sacrificing important parts of the frame. In practice, this means you can edit a single master video and instantly publish it across YouTube, Reels, TikTok, and Shorts without creating separate versions for each platform.

Free PDF Parser A free, open-source PDF parser has just appeared on GitHub that converts documents into formats optimized for large language models. OpenDataLoader PDF transforms any PDF into Markdown, JSON, or HTML, making it easy to use in RAG pipelines, AI agents, search systems, and LLM-powered applications. It also features a hybrid mode that uses OCR for scanned PDFs where the text is embedded as images, while accurately extracting content from complex tables, formulas, and diagrams. The best AI on AIMarket 🚀

First Look at Seedance 2.5 The first footage from Seedance 2.5 has surfaced. The upcoming Ultra-Long Video (Beta) mode, powered by Seedance 2.5, has also been used to generate a three-minute video. The feature is currently available only in a closed beta, but the developers have confirmed that a public release is just around the corner. The announced capabilities match what the team had previously previewed. The best AI on AIMarket 🚀

Black & White Photo Create a highly stylish black-and-white photoshoot using the following prompt:

Preserve the person’s appearance 100%. Do not alter facial features or proportions. Shot with a professional camera, close-up portrait, black-and-white photo, hair exactly as in the reference photo. The subject looks directly into the camera with a playful expression. Slightly underexposed. Head tilted toward the shoulder while maintaining eye contact with the camera. Wearing a black cotton shirt with the top buttons undone. Light-colored wall background. Sunlight partially falls across the eyes. Hyperrealism.

Google AI Studio has added support for custom URLs for apps built with the platform Now, when you publish an app created in AI Studio, you can choose a clean, memorable URL instead of the default long and complex link. For example, instead of a random address, you can use something like atelier.ai.studio, where the name reflects your project or service. Google says this feature allows AI Studio apps to have unique, branded web addresses, making them feel more like full-fledged web services. The best AI on AIMarket 🚀

🚀 OpenAI introduced the GPT-5.6 family, featuring three models: Sol, Terra, and Luna. Sol, the flagship model, scored 53.6 on Agents’ Last Exam—13.1 points ahead of Claude Fable 5. At the medium reasoning level, it still leads by 11.4 points while costing about 4× less. Terra and Luna also outperform Claude Fable 5, with inference costs up to 16× lower. On the Artificial Analysis Coding Agent benchmark, Sol scored 80.0 versus 77.2 for Claude Fable 5, using less than half the tokens, running 2× faster, and costing one-third less. GPT-5.6 also introduces Ultra mode, where multiple AI agents work in parallel to deliver faster results. It can now review rendered output, detect visual issues, and automatically refine them instead of only generating code. Artifacts have been upgraded as well, producing cleaner presentations, documents, and spreadsheets with template support and export to common tools. ChatGPT Work begins rolling out today on web and mobile for Pro, Enterprise, and Edu users.

Higgsfield Launches Apps You can now build fully functional applications with any Higgsfield image and video generation models built in. There’s no need to connect or configure anything separately—everything works seamlessly within a single system. Development is powered by Fable 5, a model that writes code at the level of a senior software engineer. It handles the entire technical side while users simply describe what they want in plain language. With Apps, you can create beautifully designed websites of any complexity, browser extensions, plugins, desktop applications, and mobile apps. Once your project is ready, you can deploy it with a single click. Apps is available in Supercomputer and in Claude via Higgsfield MCP.

Zuckerberg enters the game: Meta AI launches Muse Image Muse Image is the first image generation model developed by Meta’s own MSL research lab. Instead of simply creating images from a text prompt, it works as an AI agent. Before generating an image, it analyzes the prompt, searches the web for relevant context, and builds a plan for what the final result should look like. As a result, users are much more likely to get exactly what they had in mind on the very first try. It also supports multi-reference composition, allowing multiple source images to be blended into a single, coherent generation. On top of that, it offers iterative editing, so users can refine an image step by step without losing its style or having to start over from scratch. Muse Image is also designed to be highly personalized. Users can upload their own photos, tag friends with the @ symbol, and even pull inspiration from what’s currently trending.

Tencent has open-sourced Hy3 Hy3 is Tencent’s flagship 295-billion-parameter model, outperforming competitors that are 2–5× larger. The model uses a Mixture-of-Experts (MoE) architecture. Although it has 295 billion parameters in total, only 21 billion are active at any given time, making it faster and more memory-efficient than dense models of a similar size. The model contains 192 experts, with eight activated per request, and supports a 256K-token context window. Tencent focused primarily on agentic workflows and production reliability. The hallucination rate dropped from 12.5% to 5.4%, while errors on everyday factual knowledge were reduced from 25.4% to 12.7%. On the MRCR benchmark for long, multi-turn conversations, the score jumped from 42.9% to 75.1%. Tencent also conducted a blind evaluation involving 270 experts performing real-world work tasks. Hy3 scored 2.67 out of 4, compared to 2.51 for GLM-5.1.

LongCat-2.0 is now fully open source with no restrictions Meituan has released the weights and inference code for its trillion-parameter LongCat-2.0 model under the MIT License. No restrictions: you can use it commercially, modify it, and integrate it into your own products. 1.6 trillion parameters in total, with about 48 billion parameters activated per token using a Mixture of Experts (MoE) architecture. 1 million-token context window out of the box. Optimized for agentic workflows, with native integration for Claude Code, OpenClaw, and Hermes Agent. As a reminder, this is the same model that spent two months in the OpenRouter top 3 under the codename Owl Alpha, with developers using it without knowing who was behind it. Now you can download everything and run it locally. The model weights are available on Hugging Face, and the inference code is on GitHub. The best AI on AIMarket 🚀

Web designers, here’s a gem: a massive library of backgrounds, gradients, and shaders for websites, apps, and presentations is now completely free. The creators of Pad Shaders have open-sourced the entire project. According to the developers, all plugins, tools, and other resources are now free to use in both personal and commercial projects. Get access here. The best AI on AIMarket 🚀

Wan Video introduces Music to Dance Upload a character image and a music track, and the model generates a dance video with choreography synchronized to the beat. It supports five dance styles: street, tap, Latin, K-pop, and Chinese classical. The feature is built on top of the native audio-video synchronization technology Alibaba introduced in the Wan 2.5 series. That version already allowed users to generate videos with synchronized speech, sound effects, and music from arbitrary prompts. Music to Dance turns that capability into a dedicated, ready-to-use mode with selectable dance styles, rather than relying on general prompt-based generation. The best AI on AIMarket 🚀

Two main ways to get Seedance to generate exactly what you want. 1️⃣ References and storyboards created in ChatGPT 👉 Seedance 2️⃣ A scene you animated in Blender, plus reference images 👉 Seedance The first one is obviously easier but which result do you like better? The best AI on AIMarket 🚀