In the days leading up to Christmas, Anthropic has quietly released significant updates to Claude Code CLI in release versions 2.0.70 to 2.0.74 that include many incremental quality improvements and bug-fixes to improve interactivity and developer experience in Claude Code. These include:
Native Language Server Protocol (LSP) support, shifting how the AI agent navigates codebases from text-based searching to protocol queries, which helps to instantly find code definitions and references while drastically cutting token usage on refactoring tasks.
Added Claude in Chrome (Beta) feature that works with the Chrome extension to let you control your browser directly from Claude Code.
Terminal UI enhancements like search filtering in the plugin discovery screen, better clipboard history navigation (alt-y yank-pop), and clickable links to open attached images in the default viewer.
Numerous bug-fixes, such as fixing Claude Code on Windows rendering issues, reducing flickering issues, and addressing model service issues with Opus or Sonnet.
Each one by itself is minor, but combined with Claude Opus 4.5 AI model the updates make Claude Code a more powerful AI coding agent than ever. A Claude AI sub-Reddit comment summary explains Claude Code’s position:
Despite the bug, there’s strong agreement that Claude Code is significantly better than competitors like Cursor, especially for large codebases, due to its superior prompting and tool use without relying on flawed indexing.
Claude Code has reshaped how we do AI coding, spawned imitators like OpenAI’s Codex and Gemini CLI, and has served as a great platform for general agentic AI, with uses beyond coding. Anthropic is continuing to keep Claude Code ahead of competition, which makes it the Top AI Tool not just for this week, but for 2025.
Chinese AI company Z.ai released GLM-4.7, an open-source frontier AI model with excellent performance on coding and reasoning. Designed for production workflows and reliable code generation , it offers stronger long-context understanding (on 200K token context length) and does well on end-to-end multi-step agentic tasks. Tool use is integrated with Z.ai’s full-stack tools. Benchmark results are stellar: GLM-4.7 gets 73.8% on SWE-bench Verified and scores 68 on Artificial Analysis’ Intelligence Index, beating all other Chinese AI models and Claude Sonnet 4.5, just behind the top 3 frontier AI models. It’s the latest SOTA open source AI model.
GLM-4.7 has text only for input and output, so it is suitable for writing and coding but not multi-modal tasks. Pricing for GLM-4.7 API is $0.6/M token input and $2.20/M token output, in the range of the cost of Gemini 3 Flash.
Chinese startup MiniMax released MiniMax M2.1, an open-weights AI model that emphasizes agentic capabilities for complex tasks with improved interleaved thinking. M2.1 shows significant improvements over M2 on real-world performance, with improved coding ability across multiple programming languages and more concise and efficient responses. Aside from API access, MiniMax offers a coding plan that runs a low as $10/month as well as an AI agent interface.
On December 22, OpenAI rolled out a personalized annual review feature “Your Year with ChatGPT,” a feature similar to Spotify Wrapped. Available to free, Plus, and Pro users in the US, Canada, UK, Australia, and New Zealand, the feature provides personalized awards based on usage patterns, a custom poem, and an AI-generated image reflecting the user’s interests throughout the year.
OpenAI added adjustable personality controls to ChatGPT which allow users to control ChatGPT’s conversational tone. Characteristics such as warmth and enthusiasm can be adjusted directly within the interface, enabling users to tailor AI behavior to professional, creative, or casual contexts. The update responds to user response to GPT-5 lacking the ‘warmth’ of GPT-4o and reflects the increasing levels of personalization in AI products.
OpenAI has released a security update for its ChatGPT Atlas enterprise suite, specifically focusing on hardening the system against prompt injection attacks. Open AI acknowledged that prompt injection attacks on AI browsers like ChatGPT Atlas are unlikely to ever be fully solved, and that mitigations will be ongoing:
We view prompt injection as a long-term AI security challenge, and we’ll need to continuously strengthen our defenses against it (much like ever-evolving online scams that target humans).
OpenAI is taking an approach they hope will automatically discover novel attack strategies internally before they can be exploited against AI users in production.
OpenAI in implementing their own support for Skills, the capability first introduced by Anthropic for Claude in October. They also introduced rich text editing, in the Formatting Module, for ChatGPT that enables automatic formatting of generated content into professional document formats, such as Word.
LG is unveiling CLOiD, a new AI-powered domestic robot built to perform everyday household tasks using two articulated arms with five-finger dexterity. The consumer home robot incorporates LG’s “Affectionate Intelligence” system, intended to support natural interaction, contextual awareness, and adaptive task execution. LG plans to demonstrate CLOiD publicly at CES 2026.
AI slop hits science: A Cornell University study found that AI writing tools like ChatGPT are helping researchers publish up to 50% more papers, but the quality of papers is slipping:
“But the growing volume of AI written text is also making it harder for key decision makers to tell meaningful work apart from low value content.”
Researchers warn that many AI-polished papers fail to deliver real scientific value, creating a growing gap between slick writing and meaningful results that complicates peer review and funding decisions.
OpenAI published the paper Monitoring Monitorability on how to monitor chain-of-thought reasoning in AI models. The study contributes to understanding how AI systems can be made more transparent and auditable through different evaluation types (intervention, process, and outcome-property) and introduce a monitorability metric and a broad evaluation suite. Monitoring AI reasoning is important for AI safety as more capable AI models are deployed in sensitive applications.
Nvidia announced a non-exclusive licensing agreement with AI chip startup Groq valued at approximately $20 billion, in Nvidia’s largest deal ever that amounts to an acquisition in all but name. Groq founder Jonathan Ross and other key executives will join Nvidia to help integrate Groq’s Language Processing Unit (LPU) technology, allowing Nvidia to integrate Groq’s inference architecture into its broader ecosystem, while Groq will continue to operate independently to support its Groq cloud service.
Wall Street analysts view this as a strategic defensive move by Nvidia to secure its dominance in the AI inference market, absorbing competitive technology and strengthening its AI hardware roadmap.
Samsung announced plans for a standalone AI ecosystem exhibit at CES 2026. The exhibition at CES will feature product demonstrations, immersive experiences, and technical forums highlighting how AI integrates across Samsung devices and services. Samsung is positioning their AI suite as a platform-level capability rather than a single product line.
American Airlines uses an AI system called Connect Assist to manage tight flight connections, using it to determine when flights should be held briefly for passengers with close connections. The system balances operational efficiency with customer experience by modeling downstream impacts in real time. American has been using AI to improve operations and decision-making.
Pentagon will add xAI’s Grok to its AI service in early 2026. This allows all military and civilian personnel to use xAI’s Grok models in addition to the already-approved Gemini models for enterprise AI use.
Databricks announced a $4 billion funding round at a $134 billion valuation. The company surpassed $4.8 billion in revenue run-rate with over 55% year-over-year growth. They will use the funds to develop AI agent platforms including Lakebase, Agent Bricks, and Databricks Apps.
AI scientist Yann LeCun confirmed the launch of Advanced Machine Intelligence (AMI) Labs, an AI startup focused on developing AI world models to understand the physical world rather than just generating text. LeCun will serve as executive chairman and Nabla co-founder Alex LeBrun will serve as CEO. The news was shared on LinkedIn.
OpenAI is in talks to raise up to $100 billion in funding that could value the company at up to $830 billion. The company aims to close the round by the end of Q1 2026 to support its commitments to invest heavily in compute infrastructure and development.
AI coding assistant Cursor acquired automated code review startup Graphite, extending Cursor’s platform beyond code generation into debugging and pull-request management. The deal follows Graphite’s recent $52 million Series B funding. Cursor completed multiple acquisitions in recent months to consolidate their lead in AI-native developer tools.
AI drove the stock market higher in 2025, and a market analysis report suggests the Nasdaq Composite’s 24% 2025 return marks a shift in the AI sector from speculative hype to a “Monetization Era.” The report suggests that the gains in 2025 are driven by tangible revenue growth from companies successfully deploying AI into production workflows. Investors are not rewarding mere roadmaps but are capital allocation capital toward firms demonstrating “hard revenue” from AI integrations.
New York passed the RAISE Act to regulate AI frontier AI models, establishing requirements for AI frontier model developers to create and publish safety protocols and report incidents within 72 hours.
A class-action lawsuit has been filed that alleges Adobe used pirated books to train its SlimLM AI model without permission. The complaint claims Adobe’s training data included the Books3 dataset with 191,000 copyrighted books, used via the SlimPajama dataset from Cerebras. Many other AI models have been developed on similar datasets.
In a retrospective blog post published on December 23, Google published a year-end review for 2025 highlighting research breakthroughs. Google summarized 2025 as the year AI transitioned from a tool to a utility, documenting Google’s AI model advances such as Gemini 3 and Nano Banana, as well as AI products such as the Jules and Antigravity coding tools and NotebookLM. They also highlighted impacts of AI on science, achievements in math, and the shift toward agentic systems that can think and act autonomously.
The piece serves as a well-deserved victory lap for Google, as Google has done more in AI than any other AI company in 2025, becoming the AI market leader with the broadest and best portfolio of frontier AI models and AI tools.
MIT Technology Review published its year-end ‘AI Wrapped’ feature on December 25, reflecting on another year of rapid AI development and the vocabulary that emerged to describe increasingly capable but also at times problematic AI systems. The article highlighted 14 AI terms that defined 2025. Do you know these terms?
Superintelligence. Vibe coding. Chatbot Psychosis. Reasoning. World Models. Hyperscalers. Bubble. Agentic. Distillation. Sycophancy. Slop. Physical Intelligence. Fair Use. GEO.
I’ve used most of these terms, but not GEO. GEO stands for Generative Engine Optimization, the AI era equivalent of SEO, where brands and businesses craft ways to get better exposure via AI models and chatbots.
Thanks for reading AI Changes Everything! Subscribe for free to receive new posts and support my work.
No posts
Leave a Reply