Weekly AI Rankings — June 14 – June 21, 2026

This week, z.ai released GLM-5.2, a model focused on long agent tasks. Discussions center around its ability to handle context up to 1 million tokens and comparisons with other models like GPT-5.5.

GLM-5.2 features 753 billion parameters and supports around 39 billion active parameters. The model is available on Hugging Face under the MIT license.

This week, OpenAI announced improvements in GPT-5.5 related to medical responses, following retraining with the involvement of doctors. This has drawn attention to safety and accuracy concerns regarding the model.

GPT-5.5 includes improved mechanisms for routing queries and reducing dangerous errors. The model has also faced criticism for a high frequency of hallucinations.

This week, AMD showcased a mini-PC capable of running large models, including Qwen 3-235B. This event drew attention to local inference of large models.

The device based on Ryzen AI Max+ 395 utilizes unified memory of up to 128 GB, which can be distributed between CPU and GPU. This enables efficient running of large models.

This week, Anthropic introduced Claude Opus 4.7, which was tested on NMR analysis tasks and showed comparable results to specialized software. This highlights the capabilities of general models in scientific applications.

Claude Opus 4.7 demonstrated effectiveness in molecular structure recovery, opening new opportunities for use in scientific research.

This week, a federal court dismissed xAI's lawsuit against OpenAI related to the Grok 4 model. This decision drew attention to the legal aspects of AI model development.

The lawsuit involved allegations of gathering information about the Grok 4 model through hiring competitor employees, raising ethical questions in the AI industry.

Cursor announced the Origin platform for managing AI agent code, sparking interest in new development capabilities. The acquisition of Cursor by SpaceX for $60 billion is also under discussion.

The Origin platform is set to launch in the fall and aims to manage codebases for scenarios involving mass use of AI agents.

OpenAI Codex added a Record & Replay feature for scripting on macOS this week, enhancing automation of routine tasks. There are also discussions about limits being reset for some accounts.

The Record & Replay feature allows users to record their actions and convert them into editable scripts, simplifying development.

DeepSeek launched Vision Mode in its chatbot, expanding capabilities for interacting with visual data. This event drew attention to new approaches in graph analysis and UI.

Vision Mode supports visual primitives for improved localization and spatial reasoning, which can be beneficial in various applications.

AMD showcased a mini-PC for local running of large models, drawing attention to new hardware solutions for AI. The device utilizes unified memory of up to 128 GB.

During the demonstration, Qwen3-235B was run, highlighting the capabilities of local inference.

Qwen introduced the Qwen-Robot Suite, a set of models for linking visual-language models and physical actions of robots. This event highlights advancements in robotics.

The suite includes three models: RobotNav, RobotManip, and RobotWorld, expanding the capabilities of robots to interact with their environment.