AIMarketCap | AI Trends & Prompts

US Government Blocks Anthropic’s Newest AI Models Anthropic has completely disabled access to its most advanced AI models, Fable 5 and Mythos 5, for all users worldwide. The move follows an emergency order from US authorities under export control regulations. The government reportedly required Anthropic to immediately restrict access to these technologies for all foreign nationals, both inside and outside the United States, including the company’s own international employees. Anthropic disagreed with the decision, stating that the models feature robust multilayer safety systems and that the sudden restriction undermines transparency and fairness. However, the company confirmed it is complying with the legal order. Other Anthropic AI models remain fully operational. The company is now working with government agencies to resolve the issue and restore access as quickly as possible. The best AI on AIMarket 🚀

Higgsfield has launched Higgsfield Games, a service that turns a single text prompt into a complete game With Higgsfield Games, you can describe your game in plain language, choose a genre with a single click, and the Supercomputer platform builds everything else for you. It generates the code, graphics, 3D models, and hosting. The result is a fully playable browser game with a shareable link. No installation or setup required. Multiplayer can be enabled with one button. The system automatically handles lobbies and player synchronization. 3D assets are generated directly from text. Type something like “create a low-poly island,” and the game instantly gets complete models, environments, and characters. No manual modeling or asset hunting needed. Coding is optional. But if you know how to code, you can open the generated project and modify it yourself. The entire process from prompt to playable link—takes only a few minutes. Built using Claude Fable 5 technology. The best AI on AIMarket 🚀

Real-World Comparison of Claude Fable 5 on Complex Tasks Real-world comparisons of Anthropic’s new model against Claude Opus 4.8, Gemini 3.1 Pro, and GPT 5.5 on visual tasks. The gap is enormous. The tests focused on genuinely difficult problems: 3D fluid dynamics, volumetric animation, and advanced physics simulations. In every case, Fable 5 delivers results that the other models cannot even come close to reproducing. As a reminder, Fable 5 is the most powerful model Anthropic has made available to all users. It is already accessible, and until June 22 it is free for users on the Pro, Max, Team, and Enterprise plans. The best AI on AIMarket 🚀

Gemini 3.5 Live Translate Breaks Down Language Barriers Google has introduced Gemini 3.5 Live Translate, a new audio model for instant voice translation during conversations. The system supports more than 70 languages and begins translating immediately after the user starts speaking. The model operates in real time, generating translated speech while simultaneously listening to the next part of the conversation. This eliminates awkward pauses and makes communication feel natural and seamless. Within fractions of a second, the system balances processing speed and translation quality. The AI understands context and delivers a translated audio stream at the same time. As a result, Gemini stays only a couple of seconds behind the speaker while preserving the pace, pitch, and intonation of the original voice throughout the conversation. You can try it in Google Translate on iOS and Android. The best AI on AIMarket 🚀

NotebookLM has learned to think for itself and do the work for you Google has rolled out a major update to NotebookLM. Under the hood, it now runs on new models Gemini 3.5 and Antigravity making it noticeably smarter and more capable. Each notebook now comes with its own cloud computer. NotebookLM can write code and perform complex analysis directly within the workspace, and it includes more than 100 built-in skills for a wide range of tasks. Output formats have also been significantly expanded. More than a dozen options are now available, including PDF, DOCX, Excel spreadsheets, presentations, charts, and images. Even better, these outputs remain editable after generation instead of being locked into a final file. Google Search has also been integrated. You no longer need to upload your own materials to get started NotebookLM can find the information it needs on the web by itself. For now, the update is being rolled out exclusively to Google AI Ultra subscribers. The best AI on AIMarket 🚀

Microsoft has released an agent that works for you Microsoft Scout is the first agent in a new category called Autopilots. It’s always on, works autonomously, and acts on your behalf. You don’t need to keep giving it instructions—it continuously keeps track of your tasks and priorities. Built into Microsoft 365, it can access your emails, chats, calendar, and files through Teams, Outlook, OneDrive, and SharePoint. It can schedule meetings, prepare materials, track deadlines, reserve time on your calendar, and identify stalled tasks. Over time, it learns how you work and adapts to your workflow. Scout is built on the open-source OpenClaw framework. Each agent has its own account, and sensitive actions can be configured to require manual approval. For now, Scout is only available in a private preview through Microsoft’s Frontier program. The best AI on AIMarket 🚀

Krea releases Krea 2 Turbo Krea 2 Turbo lets you create high-quality images in just 2 seconds. Compatible with various style presets, design layouts, and LoRA-style image creation tools. Try it for free on the website.

Google is testing an app that creates a personalized daily feed for you from across all Google services Dreambeans is a new experimental mobile app from Google. It connects to your Google services and compiles a personalized selection of content every day based on your interests and activity. The idea is simple: the app finds things you may have missed and surfaces content that might interest you. Everything is powered by Google’s personal intelligence technologies. For now, it is only available in the United States for Google AI Ultra subscribers aged 18 and older. Everyone else can join the waitlist on the website.

ChatGPT now truly remembers you, not just takes notes OpenAI has launched a new memory system for ChatGPT called Dreaming. The old memory, introduced in April 2024, only worked when you explicitly said "remember this." Essentially, ChatGPT kept a notebook of notes but forgot everything that wasn't written down. Over time, those notes became outdated and irrelevant. Dreaming works in the background. The system automatically analyzes your past conversations, identifies what's important, and updates memory without your commands. You don't need to ask it to remember anything — ChatGPT figures out on its own that you're vegetarian, love wildlife photography, or just got back from a trip to Singapore. Memory Summary page showing exactly what ChatGPT knows about you. You can edit, add to, or delete it. Compute costs for Dreaming have been reduced 5×. Because of this, memory will be rolled out to free users as well. Currently available for Plus and Pro in the US; everyone else will get it in the coming weeks.

Reve 2.0 — an image generation model in native 4K (16 MP), which its developers call the fastest in the world at this resolution. Images are print-ready immediately, no upscaling needed. It works in two stages: an autoregressive model generates code describing the composition, objects, style, and text, then a diffusion model renders the final image. The code makes every element addressable: you can precisely set the placement of text and objects, so text comes out sharper. Iterative editing barely accumulates artifacts — regenerating from the same code yields a stable result without degradation. Because the image is code, AI agents can read its structure and make precise edits. The model is agent-native out of the box. It debuted at #2 in Image Arena with 1280 points, losing only to GPT Image 2 and beating Gemini 3.1 Flash (Nano Banana 2).

ByteDance rolled out Bernini — a new video generation and editing system where an MLLM first "plans" the scene at a semantic level, and only then a diffusion model renders the final clip. Basically, the AI thinks first, then draws. 🤯 The project is built on WAN 2.2 and additional checkpoints from the ByteDance team, plus the models listed by the authors in the repository. It supports video generation, editing, working with references, and inserting objects into existing videos. Test it here: bernini-ai.github.io

Creating a Digital Twin in Gemini That Looks and Sounds Like You Google has added a new feature to Gemini. You can now create a digital avatar that looks and sounds like you, and use it in videos generated with Gemini Omni. You only need to record your face and voice once. After that, Gemini can use this avatar in any video based on your text prompts. Open the Gemini app. In the input field, tap the file attachment button and select Avatar. Then use your phone to scan the QR code and follow the on-screen instructions to record your face and voice. Once you’re done, return to your computer and confirm the setup. In your prompt, type the @ symbol and select your avatar. For example: “Create a video of me singing with an orchestra.” You must be at least 18 years old to create an avatar. The feature currently works only in English, but it will be rolled out more widely soon. The feature is available to Google AI Plus, Pro, and Ultra subscribers.

Odysseus: how a YouTuber became the creator of AI for everyone Famous gaming YouTuber pewdiepie went from making gaming content to developing AI. That’s how Odysseus came to be — a tool similar to Claude Cowork, with numerous additions and improvements, fully open source and available to everyone. Odysseus lets you run any local AI model on your device and turn it into a full-fledged working tool. All tasks, from planning to data processing, are performed locally under your control. Now you don’t need to connect to third-party services or pay for subscriptions to build your own AI assistant. Odysseus combines agents, memory, file and terminal management, research, and much more in one system. For local AI enthusiasts and anyone who wants to keep their data and workflows under control, Odysseus is a ready-made open-source solution that actually works right on your device. The model is here.

Codex now controls Windows: launch tasks from your phone while the computer works on its own OpenAI has added computer control support for Windows. Codex can now see the screen, click, type, and perform actions directly on your PC. Mobile support has also rolled out through ChatGPT on iOS and Android. Start a task on your computer, step away, and monitor progress from your phone. If the agent reaches a step that requires approval, you can confirm it right from the app. This is still an early version OpenAI promises further updates.

ChatGPT can generate up to 10 different image variations from a single prompt How do you do it? It’s simple: add the parameter n=○ to your prompt, replacing the circle with the number of images you want to generate (from 2 to 10). The AI will then return multiple image variations at once. Important: this feature does not work with every GPT model or every image-generation mode, but in most cases it works without any issues.

Black Forest Labs has launched a virtual clothing try-on tool. FLUX VTO is an AI-powered virtual try-on system. Upload a photo of a person and a photo of a garment, and get the result in under four seconds. Logos, stitching, prints, and hardware remain intact and look true to the original item. Users can try on up to four garments at once, transfer entire outfits or individual pieces between different models, and layered clothing—such as a shirt under a blazer—interacts correctly. Existing virtual try-on tools typically take 10–30 seconds to generate results, making them too slow for live shopping experiences and too expensive to deploy across thousands of products. FLUX VTO addresses both of these challenges. A free demo is available via the link, along with integration through FLUX MCP.

Genspark releases AI Slides 5.0 The main new feature is Slide Skills — ready-made blocks with expert logic and design applied with one click. What’s new in 5.0:

— 100 built-in skills from experts in consulting, sales, education, and other fields. Apply a skill and get their approach to structure and argumentation. — Custom skill. Upload your own presentation or PDF, describe the analysis steps, and lock in your team’s style. Everything is saved for reuse. — 5× faster. New architecture and more powerful models under the hood. — Built-in data analysis. The tool reads Excel and Google Sheets on its own and turns numbers into slides.

OpenAI reveals how they use Codex internally OpenAI has released an official PDF where engineers from ChatGPT teams share how they use Codex in their daily work. With Codex, they:

— find bugs across the entire repository — refactor dozens of files in minutes — generate tests while they sleep — merge PRs during meetings

Read the document [here], and explore Codex use cases [here].

Alibaba releases a translator that clones your voice in real time Alibaba has introduced Qwen3.5-LiveTranslate a model for simultaneous translation. It supports 3500+ language pairs, understands and writes in 60 languages, with voice output in 29. The key feature is voice cloning. The model preserves the speaker's intonation and timbre on the target language in real time, instead of switching to a synthetic voice. It can also see the screen and factor in visual context, eliminating ambiguities in translation. Custom hotwords: upload your own dictionary of brands, names, and terminology the model prioritizes them during transcription. Positioned for developers who want to embed native translation into their products without noticeable latency.

Runway Aleph 2.0 IS OUT This is the second generation of the viral model that can change objects, backgrounds, and environments in videos. It can replace almost anything from a single prompt colors, hairstyles, objects, and more. It understands scene physics and lighting while keeping the background intact. If you make a change in one frame, it applies across the entire video. You can preview edits before applying them окончательно. It generates videos in 1080p and up to 30 seconds long, which is enough for many use cases. Try it first here.