The AI world keeps evolving at record speed and this week, it’s all about personalization, open reasoning, and lightning-fast intelligence.

💬 OpenAI’s ChatGPT Evolution - Sam Altman confirms a major shift as OpenAI prepares to relax ChatGPT restrictions, introduce customizable personalities, and launch verified adult access with age-gating by December — marking a bold step toward more natural, human-like AI interaction.

🧠 Ring-1T’s Silver-Level Reasoning - InclusionAI’s open-source model achieves state-of-the-art results across reasoning benchmarks like AIME 25 and ARC-AGI-1, edging toward gold-level understanding through trillion-scale reinforcement learning on the powerful ASystem framework.

⚡ Claude Haiku 4.5 Unleashed - Anthropic’s fastest model yet brings near-Sonnet-level reasoning at triple the speed and a fraction of the cost, excelling in real-time chat, coding, and enterprise support with improved safety and 200K-token context capabilities.

From humanized conversations to open-source intelligence and lightning-speed AI, this week’s updates redefine what’s possible across every layer of artificial intelligence.

OpenAI to Relax ChatGPT Restrictions and Add Adult Features

Sam Altman announced that OpenAI will soon ease several ChatGPT restrictions after successfully improving safeguards around mental health. He explained that the restrictions had initially been put in place to prevent harm but made the experience less enjoyable for many users. A new version, expected within a few weeks, will let users customize ChatGPT’s personality — allowing it to be more expressive, friendly, humorous, or emoji-filled, similar to what people liked about GPT-4o. Altman emphasized that these changes aim to make interactions feel more natural and personal, without forcing behavior purely to boost engagement. Additionally, by December, OpenAI plans to introduce comprehensive age-gating and, following its “treat adult users like adults” principle, will permit verified adults to access more mature content, including erotica. These updates mark a shift toward giving users greater control and personalization while maintaining safety boundaries.

Open-Source Model Hits Silver-Level Reasoning

Ant Ling shared updates on the open-source reasoning model Ring-1T, highlighting its state-of-the-art performance across benchmarks like AIME 25, HMMT 25, and ARC-AGI-1, as well as solving most IMO 2025 problems with silver-level reasoning through pure language understanding. Ring-1T continues training toward gold-level reasoning capabilities. The underlying system, ASystem, is a high-performance reinforcement learning framework enabling trillion-scale training with features like unified memory pooling, direct GPU communication, and a serverless sandbox supporting over ten programming languages at 10,000 queries per second. The model, hosted on ZenMux and Hugging Face, represents InclusionAI’s push toward advanced, open-source deep reasoning.

Lightning-Fast Claude: Haiku 4.5 Unleashed

Claude Haiku 4.5, Anthropic's smallest and fastest model, is now available to all users and offers a significant leap in performance, speed, and cost efficiency. It is designed for low-latency tasks like chat assistance, coding, customer support, and multi-agent systems, delivering nearly the same coding and reasoning capabilities as larger models, but at a third of the cost and twice the speed of previous models like Claude Sonnet 4.5. The model supports a 200,000-token context window, handles text and images, and introduces features like extended thinking, computer usage, and context awareness, making it highly suitable for real-time applications. Its safety profile has been notably improved, with lower risks of concerning behaviors and alignment issues, earning a Level 2 AI Safety classification. In addition to being available via API, Amazon Bedrock, and Google Cloud’s Vertex AI, Claude Haiku 4.5 is also integrated into platforms like GitHub Copilot, offering developers more responsive and cost-effective AI assistance. Overall, Claude Haiku 4.5 represents a major step in democratizing high-performance AI for both enterprise and individual use cases, providing deep reasoning, rapid responses, and broad safety in a compact package.​

Hand Picked Video

In this video we’ll look at how to install and use Claude Code to build and ship a complete portfolio website from scratch, customize the design, and deploy it live with Vercel in minutes.

Top AI Products from this week

• AI Mode by Dappier - Conversations are the new paradigm. Get in on AI-enabled ads. AI mode by Dappier gives you Perplexity for your site, in a box. Let users get answers driven by your content, monetized with conversational agentic ads.

• Basalt Agents - Basalt, #1 AI observability tool for teams, is launching its brand new Agent Builder : prototype, test, and deploy complex AI flows composed of multiple prompts, and run them through a dataset of scenarios.

• Relace Repos - Repos is a source control product co-optimized with our suite of fast, specialized coding models. It's git compatible, high throughput, and has built-in codebase retrieval.

• Outchat AI - Create, brand, and monetize AI chats with Outchat. From setup to payments, everything you need to build your AI business in one platform.

• Snap - Snap lets you run instant usability tests with AI personas. Drop in a prototype or website, and your custom AI personas will test them like real users. Get results in minutes with transcripts, recordings, and actionable recommendations.

• Zoho SalesIQ Nova '25 - Check out Zoho SalesIQ's latest release, Nova ’25—25+ features and enhancements across Zobot, GenAI, messaging channel, workflows, integrations, and more.

This week in AI

• Nano Banana Expands Across Google Apps - Google's advanced AI image editor from Gemini 2.5 Flash, is now in Google Search and NotebookLM with 6 new styles, plus soon in Photos for easy, creative image transformation via Lens and AI Mode.

• Walmart Powers ChatGPT Checkout - Walmart partners with OpenAI for ChatGPT's Instant Checkout, enabling seamless shopping within chat. Customers can browse, plan, and buy Walmart products directly in the chat, transforming e-commerce.

• Qwen3-VL Compact Model - Qwen3-VL 4B & 8B compact models deliver lower VRAM use with full Qwen3-VL power. Outperform Gemini 2.5 Flash Lite & GPT-5 Nano, rival flagship Qwen2.5-VL-72B. FP8 versions for efficient deployment.

• Preparing for AI’s Economic Impact - Anthropic’s Economic Futures program explores policy responses to AI’s economic effects, supporting research, workforce training, tax reform, and social safety nets to prepare for diverse AI-driven scenarios.

• Google Veo 3.1 Upgrade - Google’s Veo 3.1 boosts Flow with richer audio, better narrative control, enhanced visuals, and new editing features including audio in video generation and seamless scene extensions.

Paper of The Day

The paper "SAJA: A State-Action Joint Attack Framework on Multi-Agent Deep Reinforcement Learning" proposes SAJA, a novel two-phase gradient-based attack targeting both state and action spaces in multi-agent systems. SAJA generates adversarial perturbations in states and actions sequentially, exploiting their synergistic effects to degrade cooperative multi-agent team performance more effectively than state-only or action-only attacks. Evaluated in the Multi-Agent Particle Environment, SAJA outperforms existing methods and bypasses current defenses, showcasing the need for improved robustness strategies in multi-agent deep reinforcement learning.

To read the whole paper 👉️ here